Literature
Display
← All papers

TARC: Time-Adaptive Robotic Control

Arnav Sukhija · Lenart Treven · Jin Cheng · Florian Dorfler · Stelian Coros · Andreas Krause

Preprint posted · CoRL 2026 (accepted) ↗

Robotics10 min text · CondensedOriginal paper ↗
Contents

Overview

TARC learns when to request the next policy decision as well as what action to execute, allocating more frequent decisions to demanding states.

At a glance

A reinforcement-learning policy jointly selects an action and its holding duration. A fixed query penalty encourages sparse decisions; an adaptive multiplier targets an expected discounted query budget. Experiments on a radio-controlled car, a Go1 quadruped, and simulated vision-language-action tasks reduce policy calls while preserving task performance in the tested settings. Reported inference and command-travel reductions are deployment proxies, not measured energy or wear savings. Held actions remain open-loop until the next scheduled decision.

1 Introduction

Fixed control rates must accommodate difficult moments even when most states need fewer corrections. TARC makes the decision interval state-dependent, combining task return with a query-count constraint. Its contributions are the duration-aware objective, dynamic hardware transfer, and an extension to selecting how much of a generative action chunk to execute before replanning.

Refer to caption
Figure 1: Control frequency of TARC during a perturbation experiment on the Go1. During stable standstill (a) the policy operates at 16.716.7 Hz; upon being pushed (b) it spikes up to 5050 Hz to handle the disturbance, then recovers to low frequency once the robot stabilises (c).

2 Related Work

Variable action repetition and MOSEAC [32] already trade interaction cost against performance. TARC distinguishes a fixed soft penalty from a learned multiplier targeting the rate itself. It also adapts generative policy chunk execution through diffusion steering [31]. Unlike classical event-triggered control, it learns timing from reward; it does not inherit a hand-designed stability trigger or guarantee.

3 Method

The method begins with a query-constrained control objective and expresses it as rewards at variable-length decision epochs.

timeutu_{t}queryqueryqueryqueryqueryqueryΔ​t=1\Delta t{=}1Δ​t=2\Delta t{=}2Δ​t=4\Delta t{=}4high ffmid fflow ff
Figure 2: TARC’s adaptive query schedule. The policy is queried (gray dots) only when a new action is needed; between queries, the previous action utu_{t} is held for Δ​t\Delta t steps. During demanding phases the policy issues frequent queries (short Δ​t\Delta t, red); during stable phases it holds actions longer (large Δ​t\Delta t, blue), reducing inference cost proportionally.

3.1 Problem statement

A binary indicator marks whether the policy is queried at each underlying environment step. The objective maximizes discounted task return subject to an expected discounted query count bounded by K/(1−γ). K=1 permits querying every step; lower values require skipped decisions. This normalization assumes a discount below one, and the budget is an expectation rather than a strict cap on every rollout.

Equation 1
maxπ⁡𝔼π​[∑t=0γt​r​(xt,ut)]subject to𝔼π​[∑t=0γt​qt]≤K1−γ,\max_{\pi}\;\;\mathbb{E}_{\pi}\!\left[\sum_{t=0}\gamma^{t}r(x_{t},u_{t})\right]\quad\text{subject to}\quad\mathbb{E}_{\pi}\!\left[\sum_{t=0}\gamma^{t}q_{t}\right]\leq\frac{K}{1-\gamma}, (1)

3.2 Time-Adaptive Policies

The policy chooses both a control command and an integer repetition duration, with task rewards accumulated across the held interval.

Policy Class.

The input augments state with episode time. Each action is held for one to i steps, giving a decision rate between the maximum rate divided by i and the maximum rate. Increasing i allows larger computational savings but also longer periods without a new observation-conditioned policy decision.

Equations 2–3
π⁡(st)=(ut,Δ​t),ut∈𝒰,Δ​t∈{1,…​i}\pi(s_{t})=(u_{t},\Delta t),\quad u_{t}\in\mathcal{U},\Delta t\in\{1,\ldots i\} (2)
f=fmaxΔ​t∈[fmaxi,fmax],f=\frac{f_{\max}}{\Delta t}\in\left[\frac{f_{\max}}{i},\,f_{\max}\right], (3)

Lagrangian relaxation.

Grouping the discounted return by decision epochs yields a per-decision reward equal to the within-hold discounted task rewards minus one query cost. The budget contributes a constant independent of the policy for a fixed multiplier. Consecutive decisions must be discounted by γ raised to their actual duration, as detailed in the appendix; using ordinary per-decision discounting would change the objective.

Equations 4–5
ℒ⁡(π,λ)=𝔼π​[∑tγt​r​(xt,ut)]−λ⁡(𝔼π​[∑tγt​qt]−K1−γ).\mathcal{L}(\pi,\lambda)=\mathbb{E}_{\pi}\!\left[\sum_{t}\gamma^{t}r(x_{t},u_{t})\right]-\lambda\!\left(\mathbb{E}_{\pi}\!\left[\sum_{t}\gamma^{t}q_{t}\right]-\frac{K}{1-\gamma}\right). (4)
R⁡(st,at)=(∑k=0Δ​t−1γk​r​(xt+k,ut))−c,R(s_{t},a_{t})=\left(\sum_{k=0}^{\Delta t-1}\gamma^{k}r(x_{t+k},u_{t})\right)-c, (5)

3.3 Soft and Hard Query Budget Constraints

The same objective supports either a manually selected query cost or an automatically adjusted budget penalty.

Soft constraint (fixed c ).

A fixed cost encourages fewer calls but does not enforce a particular realized rate. Its effect depends on task dynamics and reward scale, so it remains a tuning parameter. This is related to time-adaptive continuous-time reinforcement learning [29], adapted here to discrete control and chunk execution.

Hard constraint (auto-tuned λ ).

The nonnegative multiplier increases when estimated discounted query use exceeds the target and decreases otherwise. Updating it alongside the policy empirically tracks chosen budgets. The source calls this the hard-constraint variant, but finite-sample training and an expected constraint do not provide a guaranteed per-episode query cap or safety guarantee.

Equation 6
λk+1=max⁡(0,λk+αλ​(𝔼πk​[∑tγt​qt]−K1−γ)),\lambda_{k+1}=\max\!\left(0,\;\lambda_{k}+\alpha_{\lambda}\!\left(\mathbb{E}_{\pi_{k}}\!\left[\sum_{t}\gamma^{t}q_{t}\right]-\frac{K}{1-\gamma}\right)\right), (6)

4 Experimental Setup

The car performs a roughly two-meter reverse-parking maneuver with a 180-degree turn, usually involving a drift, at a maximum 30 Hz decision rate. State comes from OptiTrack, and randomized tire/motor dynamics support transfer. Episodes last about 6.6 seconds. The Go1 uses a maximum 50 Hz high-level policy and an unchanged lower-level joint controller; three twenty-second command scenarios test steady turning, changing speed, and abrupt run-to-turn transitions.

The VLA experiment uses a frozen LIBERO-fine-tuned π₀ model [4], a learned lightweight steering policy, and variable execution lengths up to fifty actions. Its main simulated task compares fixed chunks of twenty and thirty-two actions, with ten rollouts per checkpoint. In contrast to holding one motor command on the hardware tasks, the VLA executes a prefix of an action sequence. Most results use five training seeds. Unpenalized task reward and reward after a common query penalty are reported separately.

Refer to caption
(a) Radio-Controlled (RC) car.
Refer to caption
(b) Unitree Go1 quadruped.
Refer to caption
(c) LIBERO Benchmark with pick-and-place tasks.
Figure 3: Platforms used in the experimental evaluation: (a) RC car, (b) a quadrupedal robot, and (c) LIBERO benchmark.
Equation 7
r⁡(xt,at)=rstate​(xt)−w⋅‖at‖,w=0.005,r(x_{t},a_{t})=r_{\text{state}}(x_{t})-w\cdot\|a_{t}\|,\quad w=0.005, (7)

5 Results

Evaluation separates budget tracking, matched-rate comparisons, generative-policy inference, and hardware deployment. Statistical analysis treats the training seed as the random unit and uses bootstrap/permutation checks. Claimed task-reward equivalence uses two one-sided tests with a prespecified margin of 5% of baseline reward, rather than treating a nonsignificant difference as proof of equality.

5.1 Simulation Results

Simulation tests the expected query-budget constraint and whether adaptive timing offers benefits beyond simply lowering a fixed rate.

Automatic query budget enforcement.

For Go1 query ratios of 0.3, 0.4, 0.6, and 0.7, realized discounted rates track the targets with reported mean deviations below 5%. Tight budgets reduce task reward; at ratios of at least 0.6 it returns within the high-rate baseline’s confidence interval. This demonstrates a controllable trade-off, not cost-free query reduction at every budget.

Original figure
Figure 4: Automatic query budget enforcement on the Go1. (a) Realized discounted query ratio tracks the target KK with mean deviation below 5%5\% across all targets. (b) Task reward varies smoothly and predictably with KK.

Adaptive vs. fixed low-frequency control.

TARC meets the specified equivalence criterion against 50 Hz PPO while using fewer than half as many decisions. Against a fixed 20 Hz controller, its mean reward difference is +1.93 with confidence interval −0.8 to +5.4, so a mean improvement is not established. The stronger finding is consistency: per-seed rewards span 29.7–30.6 for TARC versus 21.5–31.7 for fixed 20 Hz, with a reported 12.6-fold standard-deviation ratio.

Original figure
Figure 5: Adaptive (TARC) vs. fixed-frequency control on the Go1 in simulation. Panels (a)–(b) show penalized and unpenalized reward (y-axis truncated at 20 to highlight differences); (c) shows average control frequency. TARC matches the high-frequency PPO-50 Hz baseline in unpenalized reward and attains comparable mean reward to the PPO-20 Hz baseline trained at TARC’s average operating frequency with lower variance.

Vision-language-action model.

On the main LIBERO task, average executed chunk length is 32.1±7.8, versus twenty for the baseline, reducing transformer calls by about 38% while retaining success. TARC also exceeds a fixed thirty-two-action comparison. This supports adaptive replanning within the evaluated simulated task and post-training method, not a measured 38% reduction in total robot execution time.

Original figure
Figure 6: TARC on the LIBERO LIVING_ROOM_SCENE3 environment (task 57) using π0\pi_{0} as the base controller. (a) TARC (c=20c{=}20, L∈[5,50]L\in[5,50]) matches the success rate of the fixed query-frequency baseline (QF=20\text{QF}{=}20) and exceeds the baseline querying at QF=32\text{QF}{=}32, TARC’s own average chunk length. (b) At convergence, TARC uses an average chunk length of L¯=32.1±7.8\bar{L}=32.1\pm 7.8 (range [22,43][22,43]), varying adaptively within episodes, compared to the fixed baseline’s constant L≡20L\equiv 20.

5.2 Hardware Results

Simulation-trained car and quadruped policies are deployed without hardware fine-tuning, with task reward, decision frequency, and control-command variation measured.

RC Car.

Five car rollouts per policy show similar unpenalized reward and lower variance for TARC. Total absolute throttle-command variation falls by 21%. That is a measured command-travel proxy; mechanical wear and energy consumption are not directly measured in these short experiments.

Original figure
Figure 7: RC car hardware performance, mean and standard error over 5 seeds. (a) Penalized reward (c=0.1c=0.1), (b) unpenalized reward, (c) average control frequency, (d) total motor throttle travel.

Unitree Go1.

Across three Go1 scenarios, task reward improves by 2.58 with confidence interval 0.5–5.2. The policy uses 365±2 queries per one thousand environment steps, about 64% fewer than the 50 Hz baseline. Identical per-query inference graphs make call count a useful inference-load proxy. A perturbation example switches from 16.7 to 50 Hz after detecting the disturbance, but a disturbance arriving mid-hold must wait for the next scheduled decision.

Original figure
Figure 8: Go1 hardware performance across the three evaluation scenarios. (a) Penalized reward (c=0.005c=0.005), (b) unpenalized reward, (c) average control frequency. TARC achieves the highest penalized reward and the smallest variance across scenarios.

6 Conclusion

Duration selection can reduce high-level policy inference while retaining performance across the tested tasks. The expected-budget formulation and duration-aware return provide a principled training interface; matched-frequency comparisons show the clearest adaptive benefit on the quadruped and simulated VLA tasks. Car results also show that some tasks can be handled just as well at a suitably chosen fixed low rate.

7 Limitations

One global duration applies across all controlled degrees of freedom, which may be restrictive for subsystems with different dynamics. Short hardware episodes do not measure energy savings. Safety is only implicit in task reward, and held actions cannot react until the next decision, bounded by the maximum hold duration. Explicit safety constraints and per-subsystem timing remain future work.

Acknowledgments

The work acknowledges Swiss National Science Foundation support through NCCR Automation.

Appendix A Proof of Proposition 1

The proof groups ordinary environment timesteps into variable-length decision epochs and preserves the discounted objective exactly for the stated policy class.

Setup.

A policy call occurs at the start of each epoch. Its command persists for the selected duration, and the next epoch starts after precisely that many environment steps. The query indicator is zero between epochs.

Decomposing the task reward.

The task-return sum is regrouped into within-epoch discounted rewards multiplied by the discount at each epoch’s absolute start time. Equation 8 retains the complete decomposition.

Equation 8
𝔼π​[∑tγt​r​(xt,ut)]\displaystyle\mathbb{E}_{\pi}\!\left[\sum_{t}\gamma^{t}r(x_{t},u_{t})\right] =𝔼π​[∑n∑k=0Δ​tn−1γtn+k​r​(xtn+k,utn)]\displaystyle=\mathbb{E}_{\pi}\!\left[\sum_{n}\sum_{k=0}^{\Delta t_{n}-1}\gamma^{t_{n}+k}r(x_{t_{n}+k},u_{t_{n}})\right]
=𝔼π​[∑nγtn​∑k=0Δ​tn−1γk​r​(xtn+k,utn)].\displaystyle=\mathbb{E}_{\pi}\!\left[\sum_{n}\gamma^{t_{n}}\sum_{k=0}^{\Delta t_{n}-1}\gamma^{k}r(x_{t_{n}+k},u_{t_{n}})\right]. (8)

Decomposing the constraint.

Because each epoch contains exactly one call, the discounted query count becomes a sum of discounts at epoch start times, as shown in Equation 9.

Equation 9
𝔼π​[∑tγt​qt]=𝔼π​[∑nγtn].\mathbb{E}_{\pi}\!\left[\sum_{t}\gamma^{t}q_{t}\right]=\mathbb{E}_{\pi}\!\left[\sum_{n}\gamma^{t_{n}}\right]. (9)

Combining into the Lagrangian.

Combining both sums produces the per-decision task reward minus the multiplier, plus the policy-independent budget constant. Equation 10 preserves the algebra. The resulting control process uses duration-dependent transitions and discounting.

Equation 10
ℒ⁡(π,λ)\displaystyle\mathcal{L}(\pi,\lambda) =𝔼π​[∑nγtn​(∑k=0Δ​tn−1γk​r​(xtn+k,utn)−λ)]+λ​K1−γ\displaystyle=\mathbb{E}_{\pi}\!\left[\sum_{n}\gamma^{t_{n}}\!\left(\sum_{k=0}^{\Delta t_{n}-1}\gamma^{k}r(x_{t_{n}+k},u_{t_{n}})-\lambda\right)\right]+\frac{\lambda K}{1-\gamma}
=𝔼π​[∑nγtn​R​(stn,atn)]+const,\displaystyle=\mathbb{E}_{\pi}\!\left[\sum_{n}\gamma^{t_{n}}R(s_{t_{n}},a_{t_{n}})\right]+\text{const}, (10)

Remark (training).

Training uses γ raised to the held duration in both temporal-difference residuals and generalized-advantage recursion. This semi-Markov adjustment is essential when comparing commands of different lengths.

Appendix B Experimental Details and Hyperparameters

The implementation appendix records observations, actions, dynamics randomization, rewards, network architectures, and complete training settings for both hardware platforms.

B.1 RC Car

The car setup combines delayed state/action information with randomized drifting dynamics and a position-orientation parking objective.

B.1.1 State Space

Six physical state values cover position, heading, local velocity, and angular velocity. Three recent two-dimensional actions compensate for approximately 80 ms transmission delay. The fixed baseline therefore sees twelve values; TARC adds episode time as a thirteenth.

Table 1: Components of the 13-D observation vector for the RC Car (TARC policies). The baseline uses the first 12 dimensions only.
Dimension(s) Description
0, 1 Global position (px,pyp_{x},p_{y})
2 Orientation (θ\theta)
3, 4 Local velocity (vx,vyv_{x},v_{y})
5 Angular velocity (ω\omega)
6, 7 Action at t−1t-1 (δt−1,τt−1\delta_{t-1},\tau_{t-1})
8, 9 Action at t−2t-2 (δt−2,τt−2\delta_{t-2},\tau_{t-2})
10, 11 Action at t−3t-3 (δt−3,τt−3\delta_{t-3},\tau_{t-3})
12 Episode timestep tt (TARC only)

B.1.2 Action Space

The two continuous controls are steering and throttle, normalized to −1 through 1 and rescaled to physical limits. TARC adds a duration choice to this control interface.

B.1.3 Reward Function

A smooth tolerance function combines positional and angular parking errors, peaks at the goal, and decays with distance. The main reward subtracts a small action-magnitude penalty. Equation 11 retains the exact tolerance function, avoiding conflation of task shaping with query cost.

Equation 11
rstate​(dtotal)=1(dtotal⋅s)2+1,r_{\text{state}}(d_{\text{total}})=\frac{1}{(d_{\text{total}}\cdot s)^{2}+1}, (11)

B.1.4 Simulation Dynamics

A Pacejka tire model is randomized over mass, inertia, front/rear grip, motor constant, and steering limit. Table 2 preserves the ranges that support the claimed transfer.

Table 2: Domain randomization ranges for the RC Car simulation.
Parameter Range
Mass (mm) [1.6, 1.7] kg
Inertia (IcomI_{\text{com}}) [0.03, 0.18]
Front tire grip (dfd_{f}) [0.25, 0.6]
Rear tire grip (drd_{r}) [0.15, 0.45]
Motor constant (cm​1c_{m1}) [10.0, 40.0]
Steering limit [0.4, 0.75] rad

B.1.5 Hyperparameters

Car training uses seventy-five million timesteps, 2,048 parallel environments, and five seeds. The maximum rate is 30 Hz with soft query cost 0.1; repetition limits four, five, and ten allow minima of 7.5, 6, and 3 Hz. Table 3 retains PPO settings and the five-layer policy/four-layer critic sizes.

Table 3: Training hyperparameters for the RC Car.
Parameter Value
Environment
Episode steps 200
Parallel environments 2048
Total timesteps 75×10675\times 10^{6}
Random seeds 5
TARC-specific
fmaxf_{\text{max}} 30 Hz
Switch cost cc 0.1
fminf_{\text{min}} variants {7.5,6,3}\{7.5,6,3\} Hz
PPO
Learning rate 3e-4
Batch size 1024
Minibatches 32
Updates per batch 4
Unroll length 10
Entropy cost 0.01
Discount γ\gamma 0.9
GAE λ\lambda 0.95
Clip ϵ\epsilon 0.3
Optimizer Adam
Network architecture
Policy hidden layers 5×325\times 32
Critic hidden layers 4×1284\times 128
Activation Swish

B.2 Unitree Go1

The Go1 implementation keeps the underlying joint-tracking controller and task reward fixed while changing high-level action timing.

B.2.1 Environment and Policy Architecture

MuJoCo Playground randomizes sensing and physical properties. A twelve-action policy produces joint targets tracked by an unchanged proportional–derivative controller at the underlying rate. An asymmetric critic receives privileged state during training; hardware inference uses ONNX on the onboard CPU. Thus policy-query reduction does not mean all low-level feedback or actuator updates stop.

B.2.2 Go1 Evaluation Scenarios

The three held-out command schedules probe smooth low-speed turning, varying straight-line speed, and running followed by in-place turning. Their twenty-second durations compare different responsiveness demands rather than a long-horizon energy or reliability benchmark.

Original figure
Figure 9: Go1 evaluation scenarios. Left: Gentle Curve. Center: Velocity Changes. Right: Run Then Turn.

B.2.3 Hyperparameters

Go1 training uses six hundred million timesteps, 8,192 parallel environments, five seeds, and a 50 Hz maximum policy rate. The soft query cost is 0.005; both actor and critic use widths 512, 256, and 128. Table 4 retains the full optimization configuration.

Table 4: Training hyperparameters for the Unitree Go1.
Parameter Value
Environment
Episode steps 1000
Parallel environments 8192
Total timesteps 600×106600\times 10^{6}
Random seeds 5
TARC-specific
fmaxf_{\text{max}} 50 Hz
Switch cost cc (soft constraint) 0.005
fminf_{\text{min}} variants {12.5,10,5}\{12.5,10,5\} Hz
PPO
Learning rate 3e-4
Batch size 256
Minibatches 32
Updates per batch 4
Unroll length 20
Entropy cost 0.01
Discount γ\gamma 0.97
GAE λ\lambda 0.95
Clip ϵ\epsilon 0.2
Optimizer Adam
Network architecture
Policy hidden layers [512,256,128][512,256,128]
Critic hidden layers [512,256,128][512,256,128]
Activation Swish

Appendix C Supplementary Results

The supplementary ablations vary maximum hold counts of four, five, and ten, examining both frequency savings and the consequences of longer open-loop intervals.

C.1 Simulation Results

Simulation ablations separate fixed query-cost sensitivity, adaptive-versus-fixed scheduling, and task-dependent VLA chunk lengths.

Switch Cost Sensitivity

A fixed cost changes the reward trade-off but does not directly specify the realized rate. The cost sweeps test whether practical operation follows the intended direction without destroying task performance.

Original figure
Figure 10: Switch cost sensitivity on both platforms. Left: RC car; right: Go1. Each pair shows unpenalized reward and average control frequency as a function of cc. Dashed lines indicate the theoretical lower frequency bound fmax/if_{\text{max}}/i for each TARC-ii variant.

RC Car.

For the car, each repetition-limit variant settles near a task-specific rate—approximately 10, 7.5, or 4 Hz—across the tested cost sweep. Reward remains high. Here, dynamics determine a sufficient rate more strongly than the chosen soft penalty.

Go1.

For Go1, very small penalties give little frequency reduction, intermediate penalties lower the rate while retaining reward, and costs of at least 0.05 cause reward collapse. The chosen 0.005 cost sits in the useful region; it is a platform-tuned setting rather than an automatically enforced budget.

Adaptive vs. Fixed Low-Frequency Control

Matched-frequency comparisons ask whether state-dependent allocation improves performance beyond reducing the average number of policy calls.

Go1.

All three Go1 repetition variants retain high-rate-baseline task reward at lower average rates, and their plotted means exceed the fixed 20 Hz controller. The main-text confidence interval remains the appropriate qualification on the individual mean-reward comparison, while the reduced seed variability is more clearly established.

Original figure
Figure 11: Adaptive (TARC) vs. matched fixed-frequency (PPO) control on the Go1. (a) Penalized reward (c=0.005c=0.005), (b) unpenalized reward, (c) average control frequency. Dashed lines in (c) indicate the theoretical minimum frequency of each TARC variant.

RC Car.

For the car, TARC-4 and TARC-5 perform comparably to fixed 10 and 7.5 Hz controllers. The task admits a suitable constant low rate, so adaptive timing is not necessary for a reward advantage in this comparison.

Original figure
Figure 12: Adaptive (TARC) vs. matched fixed-frequency (PPO) control on the RC car drifting task. (a) Penalized reward (c=0.1c=0.1), (b) unpenalized reward, (c) average control frequency. Dashed lines in (c) indicate the theoretical minimum frequency of each TARC variant.

Multi-Task VLA Generalization

Two additional LIBERO tasks bring the evaluation to three. Success remains about 96–100% relative to fixed twenty-action chunks, while average selected lengths differ by task. This supports task-dependent scheduling in the evaluated setup; it does not by itself establish transfer of one timing policy to arbitrary unseen tasks.

Original figure
Figure 13: TARC generalization across three libero_90 tasks. (a) Success rate: TARC matches the fixed QF=20 baseline across all tasks. (b) Mean chunk length L¯\bar{L}: TARC selects different average chunk lengths per task, reflecting task-specific replanning demand rather than a fixed rate.

C.2 Hardware Results

Hardware ablations check whether conclusions persist across repetition limits and whether simulated timing survives deployment.

Go1 Hardware Results

All three Go1 variants perform consistently across the tested command scenarios, matching or exceeding the high-rate baseline’s task reward. Figure 14 retains the per-scenario reward and frequency comparisons with five-seed uncertainty.

Original figure
Figure 14: Hardware results ablation for different ii. The different TARC variants perform consistently on the different tasks compared to the baseline. Mean and standard error over 5 seeds.

Sim-to-Real Transfer on the RC Car

Car query rates transfer closely, but TARC-10 shows a larger task-reward gap between simulation and hardware. A ten-step hold lasts up to about 330 ms, allowing model errors and disturbances to accumulate before correction. The authors suggest limits no greater than five for this fast platform; that is an empirical choice for these experiments rather than a universal bound.

Original figure
Figure 15: Simulation vs. hardware comparison for each TARC variant on the RC car. (a) Penalized reward, (b) unpenalized reward, (c) average control frequency. Frequencies transfer almost perfectly; reward gaps grow only for TARC-10, consistent with open-loop sensitivity at long repetition horizons. Mean and standard error over 5 seeds.
Concept demonstration

Open image ↗