Literature
Display
← All papers

TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation

Dongwon Son · Florian Shkurti · Jason Lee · Naman Shah · Beomjoon Kim · Dieter Fox

Preprint posted · CoRL 2026 (accepted) ↗

Robotics19 min text · CondensedOriginal paper ↗
Contents

Overview

This edition follows the complete thirty-three-page, 12 October 2026 v3, including all eight figures, ten numbered tables, thirty-nine numbered equations, three additional displayed equations and 46 references.

At a glance

TAM learns a torque correction between an existing controller and a robot with different dynamics. A long-history encoder infers a latent context; a fast adaptor combines it with current commands and recent motion. Analytic simulator labels train the module without downstream task demonstrations. One Panda model improves several real manipulation policies; transfer to other robot types is demonstrated in simulation. Torque access, a usable ideal model and sufficient actuator authority remain necessary.

1 Introduction

Dynamics mismatch can invalidate a policy even when its observations and goals remain appropriate. Randomizing policy training or identifying physical parameters addresses this problem, but can tie adaptation to a particular policy or require an identifiable parameterization. TAM instead modifies the common torque interface below independently authored policies and controllers.

Random motion histories train a task-independent local correction, reused under RL pushing, behavior-cloned flipping, MPC ball balancing and trajectory tracking. This modularity does not remove the need for coverage of relevant states, loads and actuator behavior.

Complete original figure S1.F1
Figure 1: TAM deployment interface. Independently authored policies, such as MPC ball balancing, BC cube flipping, or RL box pushing, can be paired with any compatible low-level controller, which converts their actions, including joint targets, end-effector targets, or direct torque, into nominal torque τt0\tau^{0}_{t}. TAM is inserted after this low-level-controller boundary and before the robot plant. The history encoder runs asynchronously at about 5​Hz5\,\mathrm{Hz}, summarizing the online robot-side history into a latent ztz_{t}. The torque adaptor runs at the 1​kHz1\,\mathrm{kHz} hardware control rate, reusing the most recent ztz_{t} to predict the residual torque Δ​τt\Delta\tau_{t} at every control tick. The corrected torque τt0+Δ​τt\tau^{0}_{t}+\Delta\tau_{t} is applied to the real robot with unknown joint friction/viscosity, link-mass and center-of-mass mismatch, and payload, so the robot better follows the ideal-model response under the same τ0\tau^{0} without changing the policy.

Related Work.

Appendix A compares dynamics randomization, rapid motor adaptation, identification, adaptive control and residual learning. The central distinction is where adaptation operates and how its correction is supervised.

2 Problem Definition and Scope

An opaque policy produces actions that a low-level controller converts into nominal joint torques. TAM adds a residual and clips the resulting command to target actuator limits. Ideal and target dynamics share kinematics and joint ordering; only their dynamics/actuation differ. The objective is to reproduce the ideal local acceleration under the same external generalized torque.

External torque is known only to the simulation label generator, never supplied to TAM at deployment. Inverse command maps provide teacher targets. Torque access, a usable ideal model and an invertible simulated actuator model are required. Saturation leaves residual mismatch; clipping alone establishes neither stability nor safety.

Complete original figure S2.F2
Figure 2: TAM data generation and training. Training data are task-agnostic free-space joint-motion rollouts, not downstream task demonstrations. For each rollout, we sample perturbation parameters ξ\xi and synthetic randomized joint-space waypoint references designed to cover diverse positions, velocities, torques, and dynamics mismatches. A joint impedance low-level controller tracks these waypoints, producing histories ht={q,q˙,τa}h_{t}=\{q,\dot{q},\tau^{a}\}, and we inject smooth external generalized-torque perturbations wtw_{t} to mimic the effect of object interaction. For each nominal torque τt0\tau^{0}_{t}, the ideal forward dynamics F0F_{0} provides the nominal acceleration, and the perturbed inverse command map Φξ\Phi_{\xi} computes the teacher torque τξ,t⋆\tau^{\star}_{\xi,t} needed to realize that ideal response. The history encoder maps hth_{t} to a latent ztz_{t}, the torque adaptor predicts Δ​τt\Delta\tau_{t}, and the Huber loss is applied to the teacher residual τξ,t⋆−τt0\tau^{\star}_{\xi,t}-\tau^{0}_{t}.
Equations
τt0=𝒞⁡(at,qt,q˙t)\tau^{0}_{t}=\mathcal{C}(a_{t},q_{t},\dot{q}_{t})
τtc=clip⁡(τt0+Δ​τt,τmin,τmax)\tau^{c}_{t}=\operatorname{clip}\!\left(\tau^{0}_{t}+\Delta\tau_{t},\,\tau_{\min},\,\tau_{\max}\right)
Fξ​(qt,q˙t,τtc,wt)≈F0​(qt,q˙t,τt0,wt)F_{\xi}(q_{t},\dot{q}_{t},\tau^{c}_{t},w_{t})\approx F_{0}(q_{t},\dot{q}_{t},\tau^{0}_{t},w_{t})

3 Torque Adaptation Module

Supervised local torque matching separates training from downstream tasks. The architecture separates slow inference of dynamics context from fast correction of the current command.

3.1 Data Generation and Training Objective

For each recorded joint state and sampled nominal torque, the ideal simulator computes acceleration; the perturbed inverse command map then finds the command producing that acceleration under the same external load. Equation 1 gives this teacher. An elementwise Huber loss compares corrected torque predictions and targets, averaged over joints.

Training covers random waypoint motion rather than task demonstrations. Matching these local transitions can transfer between policies, provided their state, velocity, torque and mismatch regimes are represented; it is not direct optimization of downstream success or long-horizon stability.

Equations 1
τξ⋆​(qt,q˙t,τt0,wt)=Φξ​(qt,q˙t,F0​(qt,q˙t,τt0,wt),wt),\tau^{\star}_{\xi}(q_{t},\dot{q}_{t},\tau^{0}_{t},w_{t})=\Phi_{\xi}\!\left(q_{t},\dot{q}_{t},F_{0}(q_{t},\dot{q}_{t},\tau^{0}_{t},w_{t}),w_{t}\right), (1)
ℒτ=ρδ​(τ^θ,tc−τξ,t⋆)\mathcal{L}_{\tau}=\rho_{\delta}(\hat{\tau}^{c}_{\theta,t}-\tau^{\star}_{\xi,t})

3.2 Encoder–Torque Adaptor Architecture

A latent history representation conditions a small torque network. The two modules run at different rates, allowing long context without a Transformer invocation on every hardware tick.

History encoder.

The rolling buffer contains position, velocity and applied-command history over twelve seconds. A physics feature subtracts the ideal inverse-model torque needed for estimated motion from the recorded torque. Its ideal inverse call assumes zero external load, so contact and model mismatch can both affect it; it is not a measured disturbance decomposition.

A patch-based time-series encoder compresses overlapping per-joint windows before causal attention. Earlier informative motion can remain available when recent commands are nearly constant or contact-dominated; finite history still cannot uniquely identify every possible mismatch.

Equations
etτ=τta−Φ0​(qt,q˙t,q^¨t,𝟎)e^{\tau}_{t}=\tau^{a}_{t}-\Phi_{0}(q_{t},\dot{q}_{t},\ddot{\hat{q}}_{t},\mathbf{0})
ht={qt−i,q˙t−i,τt−ia,et−iτ}i=1Kh_{t}=\left\{q_{t-i},\dot{q}_{t-i},\tau^{a}_{t-i},e^{\tau}_{t-i}\right\}_{i=1}^{K}

Torque adaptor.

Per-joint MLPs embed a short local state/torque window. Current nominal torque and the latent context condition the residual prediction through adaptive normalization, following feature-wise conditioning and adaptive layer-normalization designs. Equation 2 retains the exact input timing. The correction is added before final command clipping.

Equation 2
Δτt=Aθ(qt−L:t,q˙t−L:t,τt−L:t−1a,τt0,zt).\Delta\tau_{t}=A_{\theta}(q_{t-L:t},\dot{q}_{t-L:t},\tau^{a}_{t-L:t-1},\tau^{0}_{t},z_{t}). (2)

3.3 DAgger Fine-Tuning and TAM Residual-Torque Input

Optional DAgger-style simulation fine-tuning collects histories under TAM’s own corrected commands, reducing training/deployment distribution shift. A further variant feeds previous predicted corrections to the history encoder. This extra residual-command history is distinct from the inverse-model physics feature. Main real experiments use the original model without these refinements.

4 Experimental Results

Multi-robot pretraining uses Panda, Piper, RBY-1, iiwa14 and Flexiv Rizon4. A single Panda-simulation-finetuned model is reused across all physical experiments without DAgger, policy-specific training or real dynamics-calibration trajectories. Real hardware evidence is entirely on Panda; other robot types are simulated.

Trials normally clear history and start the task immediately. TAM passes through nominal torque until its first latent becomes available at about 0.5 seconds; that interval counts toward results. The carried-state tracking condition explicitly preserves a preceding sixteen-second motion history.

Complete original PDF block S4.F3
Figure 3: Representative real-world evaluations with direct execution (top, w/o TAM) and TAM-enabled execution (bottom). Columns show the four Panda settings used in the main experiments: RL box pushing, BC cube flipping, an MPC ball-on-plate policy, and tracking with an added payload.

4.1 Multiple Policies and Action Spaces

Table 1 reports task-specific outcomes with the same TAM weights. The ideal simulator is a reference, not a universal performance ceiling. Physical success rates have 21 pushing trials and 50 flipping trials; ball balancing has twenty runs. No statistical significance test is reported.

Complete original PDF block S4.T1
Table 1: Real-world task statistics for multiple policies. Ideal sim reports the same policy evaluated in its ideal simulation, serving as an idealized upper-bound reference. Trial counts and time limits are given in the metric column; parenthetical labels indicate deployed visual modality. Online SysID is detailed in Appendix C, and RMA variants are only defined for RL box pushing.
Original selectable table
Task Metric Reference Real robot
Ideal sim Direct Online SysID RMA TAM (w/ RMA) TAM
Box pushing (depth distill.) Success rate (2121 trials) 84.0%84.0\% 47.6%47.6\% 57.1%57.1\% 52.4%52.4\% 66.7%66.7\% 76.2%\mathbf{76.2\%}
Behavior-cloned flip (RGB distill.) Success rate (5050 trials) 73.5%73.5\% 50.0%50.0\% 34.0%34.0\% – – 72.0%\mathbf{72.0\%}
Ball-on-plate MPC Goals / 15​s15\,\mathrm{s} (2020 runs) 15.9±2.315.9\pm 2.3 8.8±2.48.8\pm 2.4 13.1±2.313.1\pm 2.3 – – 13.6±2.9\mathbf{13.6\pm 2.9}

Vision-based pushing policies.

A vision student distilled from a MuJoCo Playground-style state-based policy outputs joint torques. Real pushing succeeds in 10/21 trials directly, 12/21 with Online SysID, 11/21 with RMA, 14/21 with RMA plus TAM and 16/21 with TAM alone. Ideal simulation reports 84%. Adding TAM helps both policies, but combining RMA and TAM is not best here.

Behavior-cloned flip policy.

An ACT policy learns from scripted ideal-simulation flip demonstrations, using gripper-camera RGB and robot state to predict joint targets. Joint impedance converts them to torque before TAM. Success rises from 25/50 directly to 36/50 with TAM; SysID obtains 17/50. The ideal-simulation reference is 73.5%.

MPC ball-on-plate.

A MuJoCo MPC planner uses visual ball-state feedback and iLQR to command torques. Over fifteen seconds, mean goals increase from 8.8 ± 2.4 directly to 13.6 ± 2.9 with TAM; SysID reaches 13.1 ± 2.3 and ideal simulation 15.9 ± 2.3. TAM’s margin over SysID is only 0.5 goal relative to substantial run variation.

4.2 Tracking: Motion Alignment Under Unknown Parameters and Low-Level Controller Switches

Each real trial tracks sixteen seconds of joint-space motion, then eight seconds with operational-space control. The target segment is scored against its ideal-model reference. Reset clears adaptation at the switch; Carried retains it. Ten trials use the physical Panda, with and without an added 1 kg payload.

TAM’s reset/carried errors are 1.51/1.04 cm without payload and 2.35/1.29 cm with payload, versus Direct 4.09/6.04 cm. It beats the reported SysID and PhysPred cells. Carried results demonstrate reuse of informative prior motion, not cold-start adaptation from the first control tick.

Complete original PDF block S4.T2
Table 2: End-effector target-position tracking under two real-world physical conditions. All entries are target-position RMSE against the ideal-model reference in centimeters. “Real Panda (hidden params)” denotes the physical Franka Panda whose true physical parameters are not provided to the adaptation methods, while “Real Panda + 1 kg payload” intentionally adds an unknown 1​kg1\,\mathrm{kg} end-effector payload. For adaptive methods, Reset and Carried denote whether the online adaptation state is cleared or preserved at the start of the OSC segment; Direct has no online state and spans both columns. Values are mean ±\pm standard deviation over 1010 real-world trials per condition. Lower is better.
Original selectable table
Method Real Panda (hidden params) Real Panda + 1 kg payload
Reset Carried Reset Carried
Direct 4.09±0.394.09\pm 0.39 6.04±0.516.04\pm 0.51
Online SysID 2.67±0.632.67\pm 0.63 2.75±0.582.75\pm 0.58 4.69±0.424.69\pm 0.42 4.65±0.534.65\pm 0.53
PhysPred 1.74±0.501.74\pm 0.50 1.64±0.561.64\pm 0.56 4.15±0.414.15\pm 0.41 4.39±0.354.39\pm 0.35
TAM (ours) 1.51±0.301.51\pm 0.30 1.04±0.24\mathbf{1.04\pm 0.24} 2.35±0.392.35\pm 0.39 1.29±0.35\mathbf{1.29\pm 0.35}

4.3 Simulation Cross-Robot Generalization

Across seven simulated robots and one hundred randomized trajectories per robot, mean joint RMSE is 4.48° directly, 1.24° with SysID, 0.76° with shared TAM and 0.62° after robot-specific simulation fine-tuning. Shared weights improve both held-out types: Google Robot and Unitree Z1.

The fine-tuned column uses simulation data from each target robot and therefore is not zero-shot transfer. Physical cross-robot deployment is untested, and index-based joint correspondence assumes sufficiently similar serial-arm structure.

Complete original PDF block S4.T3
Table 3: Cross-robot tracking transfer in simulation. Values are joint-position RMSE in degrees over 100100 randomized trials; lower is better. Split marks train/test robots, and “TAM sim-ft.” is robot-wise simulation fine-tuning from the multi-robot pretrained weights. Bold/underline mark best/second-best.
Original selectable table
Robot DoF Split Direct Online SysID TAM TAM sim-ft.
Panda 77 Train 4.16±0.994.16\pm 0.99 1.43±0.531.43\pm 0.53 0.59±0.19¯\underline{0.59\pm 0.19} 0.43±0.17\mathbf{0.43\pm 0.17}
Piper 66 Train 4.88±0.964.88\pm 0.96 0.83±0.180.83\pm 0.18 0.52±0.17¯\underline{0.52\pm 0.17} 0.42±0.15\mathbf{0.42\pm 0.15}
iiwa14 77 Train 3.69±0.823.69\pm 0.82 0.99±0.350.99\pm 0.35 0.56±0.16¯\underline{0.56\pm 0.16} 0.47±0.16\mathbf{0.47\pm 0.16}
Flexiv Rizon4 77 Train 4.06±1.204.06\pm 1.20 1.33±0.511.33\pm 0.51 0.84±0.33¯\underline{0.84\pm 0.33} 0.68±0.27\mathbf{0.68\pm 0.27}
RBY-1 77 Train 3.80±0.813.80\pm 0.81 1.39±0.831.39\pm 0.83 0.84±0.16¯\underline{0.84\pm 0.16} 0.66±0.17\mathbf{0.66\pm 0.17}
Google Robot 77 Test 4.69±1.134.69\pm 1.13 1.65±0.371.65\pm 0.37 1.05±0.40¯\underline{1.05\pm 0.40} 0.93±0.32\mathbf{0.93\pm 0.32}
Unitree Z1 66 Test 6.11±1.276.11\pm 1.27 1.05±0.231.05\pm 0.23 0.93±0.19¯\underline{0.93\pm 0.19} 0.72±0.24\mathbf{0.72\pm 0.24}
Mean over robots – – 4.484.48 1.241.24 0.76¯\underline{0.76} 0.62\mathbf{0.62}

4.4 Adaptation Latency Under Abrupt Parameter Changes

Twenty simulated trajectories undergo an abrupt dynamics change. Recovery to the original model’s 1.83° steady-state threshold takes 4.6 seconds for original TAM and 1.7 seconds for both DAgger variants. Their final errors are 1.10° and 0.64° with added correction history.

Recovery includes catching up with the reference trajectory, not only latent estimation or compute time. Other reported experiments use original TAM. The separate 0.5-second startup result concerns the first latent without an abrupt change.

Complete original figure S4.F4
Figure 4: Adaptation after an abrupt dynamics change at t=0t=0 over 2020 simulated trajectories. Shading shows variation across trajectories; the dashed line marks original TAM’s steady-state RMSE. “DAgger” uses simulation fine-tuning on TAM-corrected histories; “+ TAM residual input” additionally supplies TAM’s past residual-torque outputs to the history encoder.

5 Limitations

TAM needs torque access, matched ideal/target kinematics, a usable ideal inverse model and adequate torque authority. Local supervised correction cannot guarantee long-horizon behavior, collision avoidance or constraint satisfaction. Saturated commands cannot fully compensate a load; unfamiliar contact or actuator behavior can confound history-based inference.

Appendix G gives conditional boundedness for a particular controller form, not a general guarantee for arbitrary policies, impedance controllers or OSC. Position-only interfaces require additional inversion or controller access outside the demonstrated setup.

6 Conclusion

A shared torque-level correction can improve several independently obtained policies without retraining them. The evidence supports this interface on real Panda tasks and simulated cross-robot tracking, with remaining limits from observability, startup delay, saturation, coverage and controller-specific stability conditions.

Acknowledgments

The authors acknowledge Korean IITP/MSIT, NRF, MOTIE/KEIT, InnoCORE and DARIUS support.

Appendix A Related Work

Prior work adapts simulators, policies, physical parameters or action channels. TAM uses simulation-derived local torque labels and a history-conditioned correction shared below downstream controllers.

A.1 Simulator and Actuator Adaptation

Panda identification, BayesSim, active identification and learned actuator/dynamics models can improve model fidelity. TAM instead compresses correction-relevant history without requiring a uniquely identifiable physical parameter vector. Finite excitation and incomplete observations constrain both approaches.

A.2 Policy-Side Adaptation

UPOSI, RMA, manipulator adaptation and long-context policies condition action selection on inferred context. Such context is usually learned together with the policy. TAM places adaptation downstream so existing policy internals and observation interfaces can remain independent.

A.3 Adaptive Control and Online Identification

Classical adaptive control and online parameter estimation exploit structured dynamics and excitation. Their guarantees do not automatically transfer to a learned residual. TAM’s long buffer can reuse earlier informative motion but cannot remove ambiguity between unobserved loads and actuator mismatch.

A.4 Action Transformation and Residual Learning

Grounded action transformation changes the action channel, while residual RL, ASAP and TRANSIC learn corrections with task or real-data supervision. TAM’s distinction is analytic, task-independent torque supervision in simulation and reuse beneath different controllers.

Appendix B TAM Training, Deployment, and Implementation Details

The appendix specifies randomized rollout generation, the invertible actuator model, the two-rate deployment and history preprocessing. All numerical settings are retained in Tables B.1–B.2.

B.1 Simulation Data Generation and Randomization

Training varies rigid-body, payload and actuator parameters. Table B.1 is Panda-specific: examples include ±10% body mass, ±1 cm body center of mass, payloads up to 1.5 kg and direction-dependent friction, bias, damping and dead zones. These ranges define the sampled family, not universal robustness bounds.

Rollout generation.

MuJoCo/MJX generates randomized free-space joint-waypoint rollouts under impedance control. Smooth generalized-torque perturbations over finite windows mimic contact effects; they are neither instantaneous impulses nor TAM inputs. The simulator records state, applied and nominal torque, external perturbation, time and sampled parameters.

Complete original PDF block A2.T1
Table B.1: Randomization ranges used during data generation for the Panda robot.
Original selectable table
Group Quantity Distribution / range Units / notes
Rigid-body dynamics Body mass scale 𝒰⁡[0.9,1.1]\mathcal{U}[0.9,1.1] Multiplicative
Rigid-body dynamics Body COM offset 𝒰⁡[−0.01,0.01]\mathcal{U}[-0.01,0.01] per axis m
Rigid-body dynamics Joint armature 𝒰⁡([0.01,…,0.01],[0.5,0.5,0.5,0.5,0.3,0.3,0.3])\mathcal{U}([0.01,\ldots,0.01],[0.5,0.5,0.5,0.5,0.3,0.3,0.3]) kg​m2\mathrm{kg\,m^{2}}
Payload EE payload mass 𝒰⁡[0,1.5]\mathcal{U}[0,1.5] kg
Payload EE payload local COM 𝒰⁡[−0.075,0.075]\mathcal{U}[-0.075,0.075] per axis m
Actuator Torque scale 𝒰⁡[0.99,1.01]\mathcal{U}[0.99,1.01] per direction Multiplicative
Actuator Torque bias τb+∼𝒰⁡[−1,1]\tau_{b}^{+}\sim\mathcal{U}[-1,1], τb−=τb++𝒰⁡[−0.2,0.2]\tau_{b}^{-}=\tau_{b}^{+}+\mathcal{U}[-0.2,0.2] N​m\mathrm{N\,m}
Actuator Dead zone 𝒰⁡[0,1]\mathcal{U}[0,1] per direction N​m\mathrm{N\,m}
Actuator Viscous damping 𝒰⁡[0,2]\mathcal{U}[0,2] per direction N​m​s/rad\mathrm{N\,m\,s/rad}
Nonlinear sigmoidal friction Amplitude 𝒰⁡[0.005,3.0]\mathcal{U}[0.005,3.0] per direction N​m\mathrm{N\,m}
Nonlinear sigmoidal friction Velocity width 𝒰⁡[0.02,0.2]\mathcal{U}[0.02,0.2] per direction rad/s; converted to sigmoid slope
Nonlinear sigmoidal friction Velocity shift s+∼𝒰⁡[−0.02,0.02]s^{+}\sim\mathcal{U}[-0.02,0.02], s−=−s++𝒰⁡[−0.01,0.01]s^{-}=-s^{+}+\mathcal{U}[-0.01,0.01] rad/s
History augmentation Position noise std. 𝒰⁡[10−4,10−3]\mathcal{U}[10^{-4},10^{-3}] rad
History augmentation Velocity noise std. 𝒰⁡[10−3,10−2]\mathcal{U}[10^{-3},10^{-2}] rad/s
History augmentation Velocity-history delay 𝒰⁡[0,2]\mathcal{U}[0,2] ms; integer 1​kHz1\,\mathrm{kHz} steps
History augmentation Applied-torque-history delay 𝒰⁡[0,4]\mathcal{U}[0,4] ms; integer 1​kHz1\,\mathrm{kHz} steps

Actuator perturbation model.

Equations B.1–B.8 define direction-dependent scaling, bias, damping, shifted sigmoid friction and a softened dead zone. A nonzero inner slope makes the dead-zone map invertible. The inverse first removes additive effects, then reverses the dead zone and torque scale.

Notation needs care: main Section 2 defines recorded τᵃ as the clipped command, whereas B.5 uses the same symbol for actuator-transformed effective torque. Command history and physical actuator output are distinct quantities; the printed equations are retained without inferring undocumented implementation behavior.

Equations B.1–B.8
scalej⁡(τ)={αj+​τ,τ≥0,αj−​τ,τ<0.\operatorname{scale}_{j}(\tau)=\begin{cases}\alpha_{j}^{+}\tau,&\tau\geq 0,\\ \alpha_{j}^{-}\tau,&\tau<0.\end{cases} (B.1)
biasj⁡(q˙)={bj+,q˙≥0,bj−,q˙<0,dampj⁡(q˙)={dj+​q˙,q˙≥0,dj−​q˙,q˙<0.\operatorname{bias}_{j}(\dot{q})=\begin{cases}b_{j}^{+},&\dot{q}\geq 0,\\ b_{j}^{-},&\dot{q}<0,\end{cases}\qquad\operatorname{damp}_{j}(\dot{q})=\begin{cases}d_{j}^{+}\dot{q},&\dot{q}\geq 0,\\ d_{j}^{-}\dot{q},&\dot{q}<0.\end{cases} (B.2)
fricj⁡(q˙)\displaystyle\operatorname{fric}_{j}(\dot{q}) ={Aj+​[σ⁡(kj+​(q˙+sj+))−σ⁡(kj+​sj+)],q˙≥0,Aj−​[σ⁡(kj−​(q˙+sj−))−σ⁡(kj−​sj−)],q˙<0,\displaystyle=\begin{cases}A_{j}^{+}\!\left[\sigma\!\left(k_{j}^{+}(\dot{q}+s_{j}^{+})\right)-\sigma\!\left(k_{j}^{+}s_{j}^{+}\right)\right],&\dot{q}\geq 0,\\ A_{j}^{-}\!\left[\sigma\!\left(k_{j}^{-}(\dot{q}+s_{j}^{-})\right)-\sigma\!\left(k_{j}^{-}s_{j}^{-}\right)\right],&\dot{q}<0,\end{cases} (B.3)
σ⁡(x)=11+exp⁡(−x).\displaystyle\sigma(x)=\frac{1}{1+\exp(-x)}.
dzj⁡(x)={ϵ​x,|x|≤δj​(x),sgn⁡(x)​(|x|−δj​(x)+ϵ​δj​(x)),|x|>δj​(x),ϵ=10−2.\operatorname{dz}_{j}(x)=\begin{cases}\epsilon x,&|x|\leq\delta_{j}(x),\\ \operatorname{sgn}(x)\left(|x|-\delta_{j}(x)+\epsilon\delta_{j}(x)\right),&|x|>\delta_{j}(x),\end{cases}\qquad\epsilon=10^{-2}. (B.4)
τt,ja=dzj⁡(scalej⁡(τt,j0))+biasj⁡(q˙t,j)−dampj⁡(q˙t,j)−fricj⁡(q˙t,j).\tau^{a}_{t,j}=\operatorname{dz}_{j}\!\left(\operatorname{scale}_{j}(\tau^{0}_{t,j})\right)+\operatorname{bias}_{j}(\dot{q}_{t,j})-\operatorname{damp}_{j}(\dot{q}_{t,j})-\operatorname{fric}_{j}(\dot{q}_{t,j}). (B.5)
yt,j=τt,j⋆+dampj⁡(q˙t,j)+fricj⁡(q˙t,j)−biasj⁡(q˙t,j).y_{t,j}=\tau^{\star}_{t,j}+\operatorname{damp}_{j}(\dot{q}_{t,j})+\operatorname{fric}_{j}(\dot{q}_{t,j})-\operatorname{bias}_{j}(\dot{q}_{t,j}). (B.6)
dzj−1⁡(yt,j)={yt,j/ϵ,|yt,j|≤ϵ​δj​(yt,j),sgn⁡(yt,j)​(|yt,j|−ϵ​δj​(yt,j)+δj​(yt,j)),|yt,j|>ϵ​δj​(yt,j),\operatorname{dz}_{j}^{-1}(y_{t,j})=\begin{cases}y_{t,j}/\epsilon,&|y_{t,j}|\leq\epsilon\delta_{j}(y_{t,j}),\\ \operatorname{sgn}(y_{t,j})\left(|y_{t,j}|-\epsilon\delta_{j}(y_{t,j})+\delta_{j}(y_{t,j})\right),&|y_{t,j}|>\epsilon\delta_{j}(y_{t,j}),\end{cases} (B.7)
τt,j0={zt,j/αj+,zt,j≥0,zt,j/αj−,zt,j<0.\tau^{0}_{t,j}=\begin{cases}z_{t,j}/\alpha_{j}^{+},&z_{t,j}\geq 0,\\ z_{t,j}/\alpha_{j}^{-},&z_{t,j}<0.\end{cases} (B.8)

Nominal torque query sampling.

Teacher queries add predefined Gaussian noise around each recorded rollout torque and clip to the training range. They are not commands generated by the downstream task policies. Sixty-four nominal samples per recorded state broaden local command supervision.

Equation B.9
τt0=τta+ϵt,ϵt,j∼𝒩⁡(0,στ,j2),\tau^{0}_{t}=\tau^{a}_{t}+\epsilon_{t},\qquad\epsilon_{t,j}\sim\mathcal{N}(0,\sigma_{\tau,j}^{2}), (B.9)

B.2 Deployment Details

A workstation GPU updates history context asynchronously at 5 Hz for Panda. The controller-side CPU applies the fast adaptor at 1 kHz, reusing the latest latent between updates. No external-torque estimate enters TAM. After reset, missing history is zeroed/masked and nominal torque passes through until enough samples arrive, typically 0.5 seconds; a full twelve-second buffer is unnecessary.

External torque and payload identifiability.

A filtered external-torque estimate can reveal payload or rigid-body mismatch while missing some commanded-versus-transmitted actuator effects. With both payload and actuation unknown, their contributions need not separate uniquely. TAM predicts their combined correction from robot history; neither a payload label nor external-torque measurement is supplied in the reported experiments. Such measurements remain possible future auxiliary cues.

Deployment runtime footprint.

Panda’s encoder has about 4.4 million parameters and its adaptor 63 thousand. Reported RTX 3080 footprint is 1.29 GB; encoder updates average 0.999 ± 0.045 ms over one hundred runs, while adaptor heads average 0.155 ± 0.002 ms over two hundred. These compute measurements are distinct from the deployed 5 Hz/1 kHz scheduling and closed-loop recovery latency.

B.3 Implementation Details

Table B.2 retains patch size 400, stride 200, five Transformer layers, eight heads, token width 256, latent width 64 and Huber threshold 0.01, alongside filtering, masking and deployment settings.

Complete original PDF block A2.T2
Table B.2: Primary TAM implementation hyperparameters. These values are moved out of the main method text to keep the method focused on the algorithmic structure. The cold-start entry describes online deployment timing after a reset, not a separate evaluation warm-up.
Original selectable table
Component Hyperparameter Value
History stream Nominal history sampling rate 1​kHz1\,\mathrm{kHz}
History stream Rolling history horizon 12​s12\,\mathrm{s}
History patching Patch size / stride 400400 / 200200 samples
Deployment split History encoder update rate 5​Hz5\,\mathrm{Hz}
Deployment split Torque adaptor control rate 1​kHz1\,\mathrm{kHz}
Deployment split Encoder / torque adaptor compute Workstation GPU / controller CPU
Cold start First decodable latent state ≈0.5​s\approx 0.5\,\mathrm{s} after reset
Derivative filtering Masked-fit window ±50\pm 50 samples at the nominal history rate
Derivative filtering Position / velocity equation weights 2.02.0 / 1.01.0
Derivative filtering Distance taper 1/(1+|s|)1/(1+|s|)
Derivative filtering Ridge coefficient 10−610^{-6}
History masking Idle raw-torque threshold 10−510^{-5}
History encoder Causal Transformer layers / heads 55 / 88
History encoder Internal token width 256256
History encoder Latent dimension dzd_{z} 6464
Torque supervision Nominal torque samples per state SS 6464
Torque loss Huber threshold 10−210^{-2}

Patch token construction.

Each patch flattens position, velocity, applied torque and physics residual for one joint. An MLP embeds it, learned joint IDs identify its coordinate, and causal attention uses rotary position encoding. A projection produces the per-joint latent. Equation B.10 preserves the exact patch indexing.

Equation B.10
𝒫m,j=[qℓ,j,q˙ℓ,j,τℓ,ja,eℓ,jτ]ℓ=m​R+1m​R+P.\mathcal{P}_{m,j}=\left[q_{\ell,j},\dot{q}_{\ell,j},\tau^{a}_{\ell,j},e^{\tau}_{\ell,j}\right]_{\ell=mR+1}^{mR+P}. (B.10)

Derivative filtering for the physics residual.

Acceleration comes from a masked local cubic fit jointly using position and velocity equations, with stronger weight on nearby observations and small curvature regularization. The ±50-sample window uses the available recorded history. Invalid/idle rows are masked; insufficient data triggers interpolation or filtered velocity differences, then zero acceleration. The same procedure reconstructs hardware history.

Closed-loop rollout loss and training-time augmentations.

A short closed-loop loss rolls the perturbed simulator under sampled TAM-corrected commands, updates the local window and reapplies teacher supervision. Long-context latents still come from logged history. Training also injects state noise and randomized velocity/torque-history delays to reduce sensitivity to real observation artifacts.

Appendix C Torque Retargeting and Explicit-Parameter Baselines

The analytic inverse map generates training commands and is reused by explicit-parameter baselines. This shared retargeting interface separates parameter estimation quality from the commanded ideal acceleration.

C.1 Inverse Torque-Command Map

Equation C.1 applies rigid-body inverse dynamics, subtracts the known external generalized torque and inverts the modeled actuator effects. Keeping the same external load in ideal forward and target inverse dynamics preserves its intended physical effect instead of silently canceling contact.

That preservation is structural in the simulator teacher, not separately certified for the learned network. Label generation removes simulator torque limits and softens/clips problematic inverse branches; hardware limits are applied later, so some ideal targets may be unreachable.

Equation C.1
Φξ​(qt,q˙t,q¨t,wt)=Bξ†​(IDξ⁡(qt,q˙t,q¨t)−wt,qt,q˙t).\Phi_{\xi}(q_{t},\dot{q}_{t},\ddot{q}_{t},w_{t})=B_{\xi}^{\dagger}\!\left(\operatorname{ID}_{\xi}(q_{t},\dot{q}_{t},\ddot{q}_{t})-w_{t},q_{t},\dot{q}_{t}\right). (C.1)

C.2 Online SysID and PhysPred Baselines

Online SysID uses a particle-filter physical estimate; PhysPred uses a history network to predict explicit parameters, inspired by Dynamics as Prompts. Both replace TAM’s direct latent-to-residual mapping with the same analytic ideal-acceleration retargeting. PhysPred is a controlled baseline here, not a reproduction of every detail of that prior method.

Online SysID parameter estimation.

The filter maintains 512 parameter particles and a fifty-sample buffer, scoring 49 one-step transitions with weighted position/velocity errors. It updates every fifty samples, resamples below half the effective population and adds bounded jitter. GPU MJWarp evaluates candidate dynamics outside the hard real-time torque loop.

It sees the same robot-side signals, ideal model and torque limits as TAM, without rewards or ground-truth parameters. It additionally receives a hand-chosen parameter family, bounds and payload-body index. Non-smooth, multimodal actuator effects motivate a particle filter rather than a local Gaussian approximation.

Equations C.2–C.3
𝒟t={(qi,q˙i,τia,qi+1,q˙i+1,mi)}i=t−Nt−1,\mathcal{D}_{t}=\left\{(q_{i},\dot{q}_{i},\tau^{a}_{i},q_{i+1},\dot{q}_{i+1},m_{i})\right\}_{i=t-N}^{t-1}, (C.2)
ℓi​(ξ(k))=‖[wq​(q^i+1​(ξ(k))−qi+1)wq˙​(q˙^i+1​(ξ(k))−q˙i+1)]‖22,\ell_{i}\!\left(\xi^{(k)}\right)=\left\|\begin{bmatrix}\sqrt{w_{q}}\,(\hat{q}_{i+1}(\xi^{(k)})-q_{i+1})\\ \sqrt{w_{\dot{q}}}\,(\hat{\dot{q}}_{i+1}(\xi^{(k)})-\dot{q}_{i+1})\end{bmatrix}\right\|_{2}^{2}, (C.3)

Shared torque retargeting.

The estimated model inverts the ideal acceleration induced by the current nominal command, with zero external-torque input. The resulting correction is clipped to the same joint limits. This does not eliminate finite-window ambiguity, parameter-family misspecification or insufficient torque authority.

Equations C.4–C.5
q¨t0=F0​(qt,q˙t,τt0,0).\ddot{q}^{0}_{t}=F_{0}(q_{t},\dot{q}_{t},\tau^{0}_{t},0). (C.4)
τtsysid=Φξ^t​(qt,q˙t,q¨t0,0),Δ​τtsysid=τtsysid−τt0.\tau^{\mathrm{sysid}}_{t}=\Phi_{\hat{\xi}_{t}}(q_{t},\dot{q}_{t},\ddot{q}^{0}_{t},0),\qquad\Delta\tau^{\mathrm{sysid}}_{t}=\tau^{\mathrm{sysid}}_{t}-\tau^{0}_{t}. (C.5)

Appendix D Tracking Experiment Details

Table D.1 distinguishes task success from tracking error and separates policy type from the downstream controller. The following protocols explain both carried-state reuse and simulated transfer.

Complete original PDF block A4.T1
Table D.1: Experimental taxonomy. This table separates policy type, low-level controller, and metric to avoid conflating tracking, behavior-cloned contact policies, and task execution.
Original selectable table
Experiment Policy type Low-level controller Metric
Tracking Reference tracking with unknown parameters Joint impedance / OSC Joint or end-effector tracking error
Vision pushing RL policy Direct torque Success rate
Behavior-cloned flip BC policy Joint impedance Success rate
Ball-on-plate MPC policy Direct torque Goals per run

D.1 Source-to-OSC Evaluation Protocol

All methods share sixteen-second source joint trajectories and eight-second OSC targets. The source uses randomized sinusoidal amplitude, period and phase; the target uses five Cartesian waypoints. The paper reports target-position RMSE, while logging source tracking and terminal error separately.

Reset conditions clear adaptation at the switch; Carried retains it. Figure D.1 plots one carried-state payload trajectory and does not show all trials. Exact gains, waypoint ranges and sampling intervals remain in the original source.

Complete original figure A4.F1
Figure D.1: Representative carried-state OSC target trajectories for the real Panda with an added 1​kg1\,\mathrm{kg} payload, corresponding to the payload setting in Table 2. The plots show end-effector xx, yy, and zz position coordinates over the 8​s8\,\mathrm{s} OSC segment for the ideal-model reference, Direct, TAM, Online SysID, and PhysPred. Only carried-state adaptive results are shown.

D.2 Cross-Robot Tracking Protocol

Each simulated robot uses its own MuJoCo model and joint defaults. One hundred randomized sixteen-second joint references yield mean ± standard deviation. Shared pretraining excludes Google Robot and Unitree Z1; robot-wise fine-tuning then uses target simulation without new real data.

D.2.1 Cross-Robot TAM Transfer Implementation

Robots use their own joint coordinates and the first N entries of a learned joint-ID table. Per-joint embeddings, attention and adaptor weights are shared; no learned inter-robot joint mapping is added. Target models supply normalization, ideal-model residuals and torque limits.

This assumes roughly comparable ordered serial chains, not semantic joint matching or arbitrary morphology transfer. Different ordering/actuators can degrade performance; more joints than the configured ID table require extension and training.

D.2.2 Cross-Robot Physics-Residual Ablation

Removing the physics residual raises mean simulated joint error from 0.76° to 1.64°. Degradation is particularly large on held-out Google Robot, 1.05°→3.58°, and Unitree Z1, 0.93°→2.43°. Every row worsens, though Piper changes only 0.52°→0.54°. The feature helps within this robot family; it does not establish morphology-independent invariance.

Complete original PDF block A4.T2
Table D.2: Extended version of Table 3 with the no-eτe^{\tau} ablation added. All values from Table 3 are repeated for comparison. Values are joint-position tracking RMSE in degrees over 100100 randomized simulation trials; lower is better.
Original selectable table
Robot DoF Split Direct Online SysID TAM w/o eτe^{\tau} TAM TAM sim-ft.
Panda 77 Train 4.16±0.994.16\pm 0.99 1.43±0.531.43\pm 0.53 1.27±0.321.27\pm 0.32 0.59±0.19¯\underline{0.59\pm 0.19} 0.43±0.17\mathbf{0.43\pm 0.17}
Piper 66 Train 4.88±0.964.88\pm 0.96 0.83±0.180.83\pm 0.18 0.54±0.140.54\pm 0.14 0.52±0.17¯\underline{0.52\pm 0.17} 0.42±0.15\mathbf{0.42\pm 0.15}
iiwa14 77 Train 3.69±0.823.69\pm 0.82 0.99±0.350.99\pm 0.35 0.94±0.230.94\pm 0.23 0.56±0.16¯\underline{0.56\pm 0.16} 0.47±0.16\mathbf{0.47\pm 0.16}
Flexiv Rizon4 77 Train 4.06±1.204.06\pm 1.20 1.33±0.511.33\pm 0.51 1.43±0.741.43\pm 0.74 0.84±0.33¯\underline{0.84\pm 0.33} 0.68±0.27\mathbf{0.68\pm 0.27}
RBY-1 77 Train 3.80±0.813.80\pm 0.81 1.39±0.831.39\pm 0.83 1.30±0.231.30\pm 0.23 0.84±0.16¯\underline{0.84\pm 0.16} 0.66±0.17\mathbf{0.66\pm 0.17}
Google Robot 77 Test 4.69±1.134.69\pm 1.13 1.65±0.371.65\pm 0.37 3.58±1.153.58\pm 1.15 1.05±0.40¯\underline{1.05\pm 0.40} 0.93±0.32\mathbf{0.93\pm 0.32}
Unitree Z1 66 Test 6.11±1.276.11\pm 1.27 1.05±0.231.05\pm 0.23 2.43±0.172.43\pm 0.17 0.93±0.19¯\underline{0.93\pm 0.19} 0.72±0.24\mathbf{0.72\pm 0.24}
Mean over robots – – 4.484.48 1.241.24 1.641.64 0.76¯\underline{0.76} 0.62\mathbf{0.62}

Appendix E Policy Implementation Details

Policy-specific perception and planning remain outside TAM. Their details matter for interpreting the real experiments, especially what “no real calibration” means: no real dynamics-calibration trajectories are used for TAM, while task vision still uses geometric camera/plate calibration.

E.1 Vision-Based Box Pushing

Pushing combines a privileged simulator teacher, a point-cloud student and a separate real object-pose estimator used for state estimation and success measurement.

Policy training.

A PPO teacher observes simulator joint state, end-effector/object poses, goal and previous action. Reward encourages side contact, position/orientation alignment and smooth robot actions. The student is subsequently distilled from this policy rather than trained directly by point-cloud RL.

Point-cloud student.

DAgger matches teacher torques from two randomized depth views. Fused points include camera-source labels, twenty synthetic gripper keypoints and the goal-box mesh. A PointNet-style encoder with mean pooling feeds a recurrent GRU alongside proprioception/goal inputs; an auxiliary box-vertex head encourages object geometry retention.

Real-world box-pose estimation and point-cloud preprocessing.

Calibrated camera transforms project depth into robot coordinates, crop the workspace and remove table points, retaining at most 6,000 points per view. A 256-particle planar box-pose filter scores the known mesh against observed points and discards pose caches older than 0.25 seconds.

Success uses this estimate: within 3 cm and 15° of the target before fifteen seconds. Thus perception accuracy contributes to the task metric; the experiment does not use simulator object state on hardware.

RMA policy-side baseline.

RMA’s privileged teacher context is distilled into a student history estimator inside the pushing policy. TAM instead operates after that policy. The RMA-trained policy is more conservative, with accumulated wrist-orientation change 250.23° versus 334.11°; this supports, without isolating causality, the authors’ explanation for RMA+TAM trailing native-policy TAM.

E.2 Behavior-Cloned Flip Policy

A scripted simulator demonstrator moves to a pre-flip pose, reorients the cube through contact and retracts. Successful episodes train an ACT policy with ResNet-18 visual features, Transformer encoder/decoder and conditional-VAE latent. It predicts chunks of seven absolute joint targets plus gripper actions from RGB and compact robot state.

TAM is added after ACT training at the impedance controller’s torque output. Physical success means the requested cube face ends on the bottom; demonstrations are synthetic, not human teleoperation.

E.3 Ball-on-Plate MPC

MPC runs at 50 Hz with a 20 ms planning timestep and 0.4-second horizon. It inserts measured arm and estimated ball states into the planner and sends the first planned torque through TAM. Goals are sampled at radial distances 2–5 cm; a 5 mm reach radius and 10 mm exit hysteresis determine events over fifteen seconds.

iLQR residual design.

A clipped proportional velocity generates a smooth local ball reference, capped at 0.04 m/s. Six residual groups penalize ball tracking, plate-center displacement, joint posture, velocity limits, torque effort and leaving the plate, with weights 40, 1, 0.1, 5, 0.2 and 1,000. These are soft planning costs, not hard safety guarantees.

Equation E.1
vkref=clip⁡(kg​(g−pkref),vmax),pk+1ref=pkref+Δ​t​vkref,v^{\mathrm{ref}}_{k}=\operatorname{clip}\!\left(k_{g}(g-p^{\mathrm{ref}}_{k}),v_{\max}\right),\qquad p^{\mathrm{ref}}_{k+1}=p^{\mathrm{ref}}_{k}+\Delta t\,v^{\mathrm{ref}}_{k}, (E.1)

Analytical ball-position estimator.

Eight startup images calibrate a checkerboard homography and camera-to-plate transform. Fixed HSV thresholds segment the red ball; a ray intersects the ball-center plane using known radius to recover planar position. Recent estimates yield velocity, and failed detections reuse the last valid position. This is an analytic visual estimator with geometric calibration, not a learned perception model.

Appendix F Additional Analyses and Ablations

Additional experiments examine architecture, latent organization, unnecessary correction on an ideal plant, unseen friction and adaptation timing. They support the tested regimes rather than unrestricted robustness.

F.1 Architecture Ablations

In one hundred simulated pushing trials, success is 21% directly, 42% with only the short-window adaptor, 67% without the physics residual, 59% with explicit parameter prediction, 70% with privileged-latent supervision and 84% for full TAM. The privileged-latent row is an alternative training target, not simply an ablated component. No uncertainty or significance analysis accompanies these success counts.

Complete original PDF block A6.T1
Table F.1: Simulation ablations for TAM architecture choices and explicit-physics baselines. Component columns mark whether each variant uses the long-horizon history encoder, torque adaptor, and physics residual feature eτe^{\tau}. The adaptation-target column distinguishes TAM’s direct torque-correction objective from explicit-parameter and privileged-latent-state baselines. Success is measured over 100100 trials. We report success rate only.
Original selectable table
Variant Component present Adaptation target Success rate
Hist. enc. Torque adaptor eτe^{\tau}
Direct transfer ×\times ×\times ×\times None 21%21\% (21/10021/100)
Short-window torque adaptor only ×\times ✓\checkmark ×\times Direct torque correction 42%42\% (42/10042/100)
No physics residual feature ✓\checkmark ✓\checkmark ×\times Direct torque correction 67%67\% (67/10067/100)
History predicts physics parameters ✓\checkmark ×\times ✓\checkmark Explicit physical parameters 59%59\% (59/10059/100)
Privileged-parameter latent-state supervision ✓\checkmark ✓\checkmark ✓\checkmark Oracle physical-parameter latent state 70%70\% (70/10070/100)
Full TAM ✓\checkmark ✓\checkmark ✓\checkmark Direct torque correction 𝟖𝟒%\mathbf{84\%} (84/10084/100)

F.2 Latent-State Bottleneck Analysis

A crossed bank replays 64 fixed trajectories under 64 physical conditions with high-gain tracking, separating motion identity from dynamics. At the same training step, latent dimensions 16/64/256 yield held-out torque RMSE 1.038/0.977/1.017. The selected width 64 is best in this small sweep, not a sharply established global optimum.

Width 256 lowers training loss but raises the trajectory-to-condition distance ratio from 0.213 to 0.341 relative to width 64. The small 16-versus-64 error gap warrants restraint; t-SNE is supporting visualization rather than proof of complete disentanglement.

Metric computation.

The evaluator encodes the first half of each replay and measures teacher-residual prediction error in the second. Latents are standardized coordinate-wise; normalized distances compare different trajectories under one condition against different conditions on one trajectory. Equations F.1–F.5 retain the teacher and ratio definitions. Lower ratio means stronger condition-based grouping under this controlled replay protocol.

Equations F.1–F.5
Δ​τξ,t⋆=Φξ​(qt,q˙t,F0​(qt,q˙t,τt0,0),0)−τt0.\Delta\tau^{\star}_{\xi,t}=\Phi_{\xi}(q_{t},\dot{q}_{t},F_{0}(q_{t},\dot{q}_{t},\tau^{0}_{t},0),0)-\tau^{0}_{t}. (F.1)
δz​(i,j)=‖z~i−z~j‖2dz.\delta_{z}(i,j)=\frac{\|\tilde{z}_{i}-\tilde{z}_{j}\|_{2}}{\sqrt{d_{z}}}. (F.2)
Dz​(same​ξ,diff​T)=𝔼c,a≠b​[δz​((c,a),(c,b))],D_{z}(\mathrm{same}\ \xi,\mathrm{diff}\ T)=\mathbb{E}_{c,\ a\neq b}\left[\delta_{z}\big((c,a),(c,b)\big)\right], (F.3)
Dz​(diff​ξ,same​T)=𝔼c≠d,a​[δz​((c,a),(d,a))].D_{z}(\mathrm{diff}\ \xi,\mathrm{same}\ T)=\mathbb{E}_{c\neq d,\ a}\left[\delta_{z}\big((c,a),(d,a)\big)\right]. (F.4)
Rz=Dz​(same​ξ,diff​T)Dz​(diff​ξ,same​T).R_{z}=\frac{D_{z}(\mathrm{same}\ \xi,\mathrm{diff}\ T)}{D_{z}(\mathrm{diff}\ \xi,\mathrm{same}\ T)}. (F.5)
Complete original PDF block A6.T2
Table F.2: Bottleneck dimension and motion-level invariance on the push-policy crossed bank (C=64C=64 physical conditions, A=64A=64 fixed trajectories). RzR_{z} is the latent-state invariance/separation ratio Dz​(same​ξ,diff​T)/Dz​(diff​ξ,same​T)D_{z}(\mathrm{same}\ \xi,\mathrm{diff}\ T)/D_{z}(\mathrm{diff}\ \xi,\mathrm{same}\ T). Lower is better for all columns. Train loss is averaged from logged evaluations around training step 150,000150{,}000; held-out τ\tau RMSE is the offline residual-torque prediction error.
Original selectable table
dzd_{z} Train loss ↓\downarrow Held-out τ\tau RMSE ↓\downarrow RzR_{z} ↓\downarrow
1616 0.8220.822 1.0381.038 0.2000.200
6464 0.7840.784 0.9770.977 0.2130.213
256256 0.7770.777 1.0171.017 0.3410.341
Complete original PDF block A6.F1
Figure F.1: Appendix t-SNE visualization of history-token embeddings for the 6464D bottleneck model. The same embeddings are colored by trajectory ID in (a) and physical condition ID in (b).

F.3 No-Mismatch Check: TAM on the Ideal Robot

On the exact ideal simulated plant, both Direct and TAM achieve 84/100 pushing successes and 15.9 ball goals per run. TAM introduces small nonzero motion differences: 0.20 ± 0.01° joint RMSE across five trajectories, with peak residual torque below 0.91 Nm. Task-level performance is unchanged in these checks, not exact identity of all motions or a universal non-degradation guarantee.

Complete original figure A6.F2
Figure F.2: Direct versus TAM execution on the exact ideal plant (zero model mismatch) for one representative 1616 s randomized source trajectory. Left column: joint positions; the Direct rollout reproduces the ideal rollout exactly, and the TAM rollout is visually indistinguishable from it. Right column: total applied torque and TAM’s residual torque; the per-joint peak residual (annotated in each panel) stays below 0.910.91 Nm, far below the joint torque limits. The dotted line marks TAM’s first latent update at t=0.5t=0.5 s.

F.4 Out-of-Distribution Friction Robustness

At twice and three times the maximum training friction, simulated pushing success is 82%/80% with TAM versus 75%/59% directly, over one hundred trials per condition. This supports robustness to those friction shifts. Real pushing/flipping also introduce contact patterns absent from random-motion training; neither result covers arbitrary loads, contacts or actuator saturation.

Complete original PDF block A6.T3
Table F.3: Box-pushing success in simulation under out-of-distribution joint friction. Friction is scaled relative to the maximum bound of the randomized training distribution; each entry uses 100100 trials.
Original selectable table
Joint friction condition TAM Direct
Ideal model (no mismatch) 84%84\% 84%84\%
2×2\times maximum training bound 𝟖𝟐%\mathbf{82\%} 75%75\%
3×3\times maximum training bound 𝟖𝟎%\mathbf{80\%} 59%59\%

F.5 Adaptation-Latency Protocol and Warm Start

Startup passes nominal torque unchanged for roughly 0.5 seconds. The source reports a 90% tracking-error reduction after this initial context becomes available without a dynamics jump. In the abrupt-change test, trajectories begin five seconds before the switch; recovery uses the same original-TAM 1.83° threshold for every variant and includes trajectory catch-up.

Appendix G Stability and Oscillation Analysis

The appendix offers a conditional tracking-error bound and a measured oscillation check. Neither establishes general safety or stability for every downstream policy/controller combination.

G.1 Bounded Tracking Errors Under a Bounded Residual

A standard manipulator Lyapunov argument treats the learned correction as a bounded input. Under the stated passivity-form controller, damping dissipates energy quadratically while bounded disturbances inject it at most linearly, yielding a finite error ceiling.

Tracking errors.

Position error e and filtered error s = ė + Λe separate the mechanical dynamics from a stable first-order position-error equation. Λ is symmetric positive definite. This tracking error e is distinct from TAM’s torque-space feature eτ.

Equations G.1–G.2
e=q−qd,e˙=q˙−q˙d,s=e˙+Λ​e,e=q-q_{d},\qquad\dot{e}=\dot{q}-\dot{q}_{d},\qquad s=\dot{e}+\Lambda e, (G.1)
e˙=−Λ​e+s.\dot{e}=-\Lambda e+s. (G.2)

Robot dynamics and controller.

The assumed base controller supplies model-based reference acceleration, Coriolis and gravity terms plus damping of s. With a compatible rigid-body model, Ṁ − 2C is skew-symmetric and Equation G.3 follows. A generic PD or OSC controller need not satisfy this form; the proof does not automatically cover every experimental controller.

Equations G.3
M⁡(q)​q¨+C⁡(q,q˙)​q˙+g⁡(q)=τ+d0.M(q)\ddot{q}+C(q,\dot{q})\dot{q}+g(q)=\tau+d_{0}.
q˙r=q˙d−Λ​e,q¨r=q¨d−Λ​e˙.\dot{q}_{r}=\dot{q}_{d}-\Lambda e,\qquad\ddot{q}_{r}=\ddot{q}_{d}-\Lambda\dot{e}.
τ=M⁡(q)​q¨r+C⁡(q,q˙)​q˙r+g⁡(q)−KD​s⏟τbase+Δ​τTAM.\tau=\underbrace{M(q)\ddot{q}_{r}+C(q,\dot{q})\dot{q}_{r}+g(q)-K_{D}s}_{\tau_{\mathrm{base}}}+\Delta\tau_{\mathrm{TAM}}.
M⁡(q)​s˙+C⁡(q,q˙)​s+KD​s=d0+Δ​τTAM.M(q)\dot{s}+C(q,\dot{q})s+K_{D}s=d_{0}+\Delta\tau_{\mathrm{TAM}}. (G.3)

Bounds and operating conditions.

Assume globally existing unique trajectories, uniformly bounded positive inertia, bounded remaining disturbances and bounded induced additive TAM torque. Final command saturation must be inactive or its discrepancy must have a separately justified bound inside the disturbance. Command clipping alone does not establish these assumptions.

Proposition G.1 bounds filtered, position and velocity errors using initial errors, inertia bounds and the smallest damping/filter eigenvalues. The full statement and formulas are preserved below.

Equations G.4–G.7
m1​IN⪯M⁡(q)⪯m2​IN,∥d0∥≤d¯0,∥Δ​τTAM∥≤τ¯Δ.m_{1}I_{N}\preceq M(q)\preceq m_{2}I_{N},\qquad\lVert d_{0}\rVert\leq\bar{d}_{0},\qquad\lVert\Delta\tau_{\mathrm{TAM}}\rVert\leq\bar{\tau}_{\Delta}. (G.4)
w=d0+Δ​τTAM,∥w∥≤d¯0+τ¯Δ=:b.w=d_{0}+\Delta\tau_{\mathrm{TAM}},\qquad\lVert w\rVert\leq\bar{d}_{0}+\bar{\tau}_{\Delta}=:b. (G.5)
Bs=m2m1​max⁡{∥s⁡(t0)∥,bλD},Be=max⁡{∥e⁡(t0)∥,BsλΛ}.B_{s}=\sqrt{\frac{m_{2}}{m_{1}}}\max\left\{\lVert s(t_{0})\rVert,\frac{b}{\lambda_{D}}\right\},\qquad B_{e}=\max\left\{\lVert e(t_{0})\rVert,\frac{B_{s}}{\lambda_{\Lambda}}\right\}. (G.6)
∥s⁡(t)∥≤Bs,∥e⁡(t)∥≤Be,∥e˙​(t)∥≤Bs+∥Λ∥​Be.\lVert s(t)\rVert\leq B_{s},\qquad\lVert e(t)\rVert\leq B_{e},\qquad\lVert\dot{e}(t)\rVert\leq B_{s}+\lVert\Lambda\rVert B_{e}. (G.7)

Step 1: Damping dominates at large error.

Differentiate kinetic-like energy ½sᵀMs. Skew symmetry cancels the inertia/Coriolis cross term, leaving damping plus input work. Cauchy–Schwarz gives negative energy derivative whenever the filtered-error norm exceeds the input bound divided by minimum damping.

Equations G.8–G.10
V⁡(s,q)=12​s⊤​M​(q)​s,12​m1​∥s∥2≤V≤12​m2​∥s∥2.V(s,q)=\tfrac{1}{2}s^{\top}M(q)s,\qquad\tfrac{1}{2}m_{1}\lVert s\rVert^{2}\leq V\leq\tfrac{1}{2}m_{2}\lVert s\rVert^{2}. (G.8)
V˙\displaystyle\dot{V} =s⊤​M​s˙+12​s⊤​M˙​s\displaystyle=s^{\top}M\dot{s}+\tfrac{1}{2}s^{\top}\dot{M}s
=−s⊤​KD​s+s⊤​w+12​s⊤​(M˙−2​C)​s\displaystyle=-s^{\top}K_{D}s+s^{\top}w+\tfrac{1}{2}s^{\top}(\dot{M}-2C)s
=−s⊤​KD​s+s⊤​w.\displaystyle=-s^{\top}K_{D}s+s^{\top}w. (G.9)
V˙≤−λD​∥s∥2+b⁡∥s∥=−∥s∥​(λD​∥s∥−b).\dot{V}\leq-\lambda_{D}\lVert s\rVert^{2}+b\lVert s\rVert=-\lVert s\rVert\bigl(\lambda_{D}\lVert s\rVert-b\bigr). (G.10)

Step 2: An energy ceiling bounds the filtered error.

The upper inertia bound converts that norm threshold to an energy threshold. Energy cannot cross above both this threshold and its initial value while its derivative is negative. The lower inertia bound then converts the energy ceiling into the stated uniform bound on s.

Equations G.11–G.13
V∗=12​m2​(bλD)2.V_{*}=\tfrac{1}{2}m_{2}\left(\frac{b}{\lambda_{D}}\right)^{2}. (G.11)
V⁡(t)≤max⁡{V⁡(t0),V∗},t≥t0.V(t)\leq\max\{V(t_{0}),V_{*}\},\qquad t\geq t_{0}. (G.12)
∥s⁡(t)∥2\displaystyle\lVert s(t)\rVert^{2} ≤2m1​max⁡{V⁡(t0),V∗}\displaystyle\leq\frac{2}{m_{1}}\max\{V(t_{0}),V_{*}\}
≤m2m1​max⁡{∥s⁡(t0)∥2,(bλD)2}=Bs2.\displaystyle\leq\frac{m_{2}}{m_{1}}\max\left\{\lVert s(t_{0})\rVert^{2},\left(\frac{b}{\lambda_{D}}\right)^{2}\right\}=B_{s}^{2}. (G.13)

Step 3: Bound position and velocity errors.

Apply the same argument to ½‖e‖² under ė = −Λe + s. Position error is bounded by the larger of its initial norm and the filtered-error bound divided by minimum filter gain. The triangle inequality then bounds velocity error.

Equations G.14–G.16
E˙=e⊤​e˙=−e⊤​Λ​e+e⊤​s≤−λΛ​∥e∥2+Bs​∥e∥.\dot{E}=e^{\top}\dot{e}=-e^{\top}\Lambda e+e^{\top}s\leq-\lambda_{\Lambda}\lVert e\rVert^{2}+B_{s}\lVert e\rVert. (G.14)
∥e⁡(t)∥≤max⁡{∥e⁡(t0)∥,BsλΛ}=Be.\lVert e(t)\rVert\leq\max\left\{\lVert e(t_{0})\rVert,\frac{B_{s}}{\lambda_{\Lambda}}\right\}=B_{e}. (G.15)
∥e˙​(t)∥≤∥s⁡(t)∥+∥Λ∥​∥e⁡(t)∥≤Bs+∥Λ∥​Be.\lVert\dot{e}(t)\rVert\leq\lVert s(t)\rVert+\lVert\Lambda\rVert\lVert e(t)\rVert\leq B_{s}+\lVert\Lambda\rVert B_{e}. (G.16)

Interpretation for TAM.

The proposition establishes bounded errors under its assumptions, without a convergence rate or settling time. More permitted residual torque can enlarge the worst-case bound; actual compensation may nevertheless reduce the combined disturbance. The result does not certify collision avoidance, contact-force limits or state constraints.

G.2 Empirical Oscillation Analysis on the Real Panda

Ten paired real-Panda rollouts compare joint-position-error spectra with and without TAM, aggregated as RMS per 1 Hz band over 0–200 Hz. Figure G.1 shows no dominant new peak, including around the encoder’s 5 Hz update. This empirical check does not rule out oscillation under other trajectories, loads or controllers.

Complete original figure A7.F1
Figure G.1: Amplitude spectrum of the joint-position tracking error on the real Panda, aggregated as RMS per 11 Hz band over ten paired rollouts of the same reference trajectories executed with and without TAM. The inset magnifies the low-frequency band around the 55 Hz latent-update rate (dotted line). No dominant joint-position-error peak appears over 00–200200 Hz, including at the 55 Hz latent-update rate.
Concept demonstration

Open image ↗