Literature
Display
← All papers

DexCompose: Reusing Dexterous Policies for Multi-Task Manipulation with a Single Hand

Dihong Huang · Zhenyu Wei · Zhuxiu Xu · Yunchao Yao · Sikai Li · Mingyu Ding

· arXiv (preprint) ↗

Robotics6 min text · CondensedOriginal paper ↗
Contents

Overview

The illustrated compositions retain a previously manipulated object while the same hand opens a door, presses a button, opens a microwave, or operates a switch.

DexCompose examples: retaining a stick, can, ball, or mug while performing a second manipulation task.
Figure 1: DexCompose composes dexterous skills through role-aware finger ownership. By separating grasp preservation from downstream interaction with asymmetric residuals, the framework reduces destructive interference and enables robust multi-skill manipulation.

At a glance

DexCompose combines two frozen dexterous policies by assigning different fingers to object retention and downstream interaction. Release tests identify candidate allocations; two learned residual controllers stabilize the held object and adapt the downstream policy.

Across sixteen simulated task pairs, composite success averages 77.43%. The base policies stay frozen, but composition is not training-free: the method learns a retention stabilizer and a task-pair-specific downstream residual.

1 Introduction

Simply switching between full-hand policies can destroy a grasp needed by the next task. DexCompose treats redundant fingers as resources: identify which fingers can be released, assign action ownership, and learn corrections within those assignments.

The framework reuses base policies across four retention skills and four interaction skills. It reduces interference in simulation; physical deployment and longer skill sequences are not demonstrated.

2 Related Work

Earlier work learns reusable dexterous controllers, sequences skills, or reserves fingers for future interactions. Residual learning adds corrections to an existing controller. DexCompose combines these ideas through explicit finger ownership and asymmetric residuals while preserving the original base-policy weights. Its diffusion base policies build on prior diffusion-policy learning. [8]

Refer to caption
Figure 2: Overview of DexCompose. Given two frozen single-task policies, we first attribute fingers according to their necessity for preserving the first task and their availability for the downstream task. The selected finger-level mask is then expanded into action-space masks. During execution, a dual-residual stabilizer preserves the first task on its assigned fingers while adapting the downstream policy on the wrist and released fingers.

3 Method

The pipeline first discovers a finger allocation, then learns controllers that respect it.

3.1 Problem Formulation

Task A establishes a physical outcome, such as a grasp. Starting from a successful Task-A state, the controller must complete Task B while retaining the object throughout. Both base policies use the same wrist-and-finger action coordinates; the composition objective requires both preservation and interaction success.

Equations 1–3
atcomp=FΘ​(ot,atA,atB).a_{t}^{\mathrm{comp}}=F_{\Theta}(o_{t},a_{t}^{A},a_{t}^{B}). (1)
τ=(s0,a0comp,s1,a1comp,…,sH).\tau=(s_{0},a_{0}^{\mathrm{comp}},s_{1},a_{1}^{\mathrm{comp}},\ldots,s_{H}). (2)
maxΘ⁡𝔼τ∼πcomp,Θ​[SA​(τ)​SB​(τ)].\max_{\Theta}\;\mathbb{E}_{\tau\sim\pi_{\mathrm{comp},\Theta}}\left[S_{A}(\tau)S_{B}(\tau)\right]. (3)

3.2 Finger Attribution

Candidate masks assign each of five fingers either to preservation or to the next task. Stored successful grasps are replayed while released fingers gradually open. Retention and clean-release rates reveal whether the object stays held without hidden support from supposedly free fingers. An LLM selects a task-aware mask from those diagnostics, considering downstream finger utility.

Equations 4–6
𝒟hold={(sihold,qiref)}i=1N.\mathcal{D}_{\mathrm{hold}}=\{(s_{i}^{\mathrm{hold}},q_{i}^{\mathrm{ref}})\}_{i=1}^{N}. (4)
qi,ttest=m¯⊙qiref+(1−m¯)⊙[(1−αt)​qiref+αt​qopen].q_{i,t}^{\mathrm{test}}=\bar{m}\odot q_{i}^{\mathrm{ref}}+(1-\bar{m})\odot\left[(1-\alpha_{t})q_{i}^{\mathrm{ref}}+\alpha_{t}q^{\mathrm{open}}\right]. (5)
m∗=AgentSelect⁡({m,Pret​(m),Pclean​(m),ηB​(m)}m∈ℳ)m^{*}=\operatorname{AgentSelect}\left(\{m,P_{\mathrm{ret}}(m),P_{\mathrm{clean}}(m),\eta_{B}(m)\}_{m\in\mathcal{M}}\right) (6)

3.3 Dual Residual Stabilizer

The preserved fingers follow the stored grasp reference plus a bounded residual. The wrist and released fingers follow the frozen downstream policy plus a separate bounded residual. Hard masks prevent either command from writing into the other’s assigned subspace. Near-zero residual initialization starts near the nominal behavior; base policies and masks remain fixed during residual learning.

Equations 7–12
Δ​qtA=βA⊙tanh⁡(πresA​(ot,m∗)),\Delta q_{t}^{A}=\beta_{A}\odot\tanh\left(\pi_{\mathrm{res}}^{A}(o_{t},m^{*})\right), (7)
qtA,pres=qref+m¯∗⊙Δ​qtA.q_{t}^{A,\mathrm{pres}}=q^{\mathrm{ref}}+\bar{m}^{*}\odot\Delta q_{t}^{A}. (8)
atA,pres=(𝟎6,qtA,pres).a_{t}^{A,\mathrm{pres}}=\left(\mathbf{0}_{6},q_{t}^{A,\mathrm{pres}}\right). (9)
Δ​atB=MB∗⊙[βB⊙tanh⁡(πresB​(ot,atB,m∗))],\Delta a_{t}^{B}=M_{B}^{*}\odot\left[\beta_{B}\odot\tanh\left(\pi_{\mathrm{res}}^{B}(o_{t},a_{t}^{B},m^{*})\right)\right], (10)
atB,exec=atB+Δ​atB.a_{t}^{B,\mathrm{exec}}=a_{t}^{B}+\Delta a_{t}^{B}. (11)
atcomp=MA∗⊙atA,pres+MB∗⊙atB,exec.a_{t}^{\mathrm{comp}}=M_{A}^{*}\odot a_{t}^{A,\mathrm{pres}}+M_{B}^{*}\odot a_{t}^{B,\mathrm{exec}}. (12)

4 Experiments

The main experiment evaluates all sixteen combinations of four retention and four interaction tasks, followed by preservation diagnostics and component ablations.

Table 1: Main comparison on 16 composite tasks. Composite success rate (%), reported as mean ±\pm standard deviation over eight seeds. “Decomp.” denotes Decomposed Action Space, “FullRes.” denotes Residual Learning, and “Ours-ZS” denotes the zero-shot DexCompose variant without the Task-B Residual.
Composite Success Rate (%)
First Task Second Task Frozen Decomp. FullRes. Ours-ZS Ours
GraspBall OpenDoor 0.00±0.000.00_{\pm 0.00} 38.94±2.4338.94_{\pm 2.43} 66.35±2.5866.35_{\pm 2.58} 73.56±6.9473.56_{\pm 6.94} 82.69±8.33
PushButton 7.69±1.927.69_{\pm 1.92} 42.31±2.1642.31_{\pm 2.16} 74.04±2.1174.04_{\pm 2.11} 76.92±4.6976.92_{\pm 4.69} 85.58±7.50
OpenMicrowave 0.00±0.000.00_{\pm 0.00} 34.62±2.5734.62_{\pm 2.57} 63.46±2.7363.46_{\pm 2.73} 64.66±7.7064.66_{\pm 7.70} 73.08±9.25
TurnOnSwitch 0.00±0.000.00_{\pm 0.00} 31.73±2.6831.73_{\pm 2.68} 55.29±3.0855.29_{\pm 3.08} 63.94±8.3363.94_{\pm 8.33} 71.88±8.93
PourMug OpenDoor 0.00±0.000.00_{\pm 0.00} 36.54±2.4736.54_{\pm 2.47} 58.65±2.8158.65_{\pm 2.81} 66.59±8.7866.59_{\pm 8.78} 75.72±8.08
PushButton 9.13±2.119.13_{\pm 2.11} 40.87±2.2440.87_{\pm 2.24} 71.63±2.1771.63_{\pm 2.17} 78.61±6.6178.61_{\pm 6.61} 85.58±7.28
OpenMicrowave 0.00±0.000.00_{\pm 0.00} 32.21±2.6132.21_{\pm 2.61} 54.81±2.9454.81_{\pm 2.94} 52.40±7.6752.40_{\pm 7.67} 62.74±8.21
TurnOnSwitch 0.00±0.000.00_{\pm 0.00} 30.29±2.7230.29_{\pm 2.72} 50.96±3.1650.96_{\pm 3.16} 55.29±7.4855.29_{\pm 7.48} 64.90±9.89
PickCan OpenDoor 0.00±0.000.00_{\pm 0.00} 35.58±2.5135.58_{\pm 2.51} 56.73±2.8456.73_{\pm 2.84} 72.84±6.6872.84_{\pm 6.68} 79.09±6.04
PushButton 8.65±2.088.65_{\pm 2.08} 39.90±2.2739.90_{\pm 2.27} 66.83±2.4666.83_{\pm 2.46} 77.64±5.0877.64_{\pm 5.08} 85.34±6.92
OpenMicrowave 0.00±0.000.00_{\pm 0.00} 31.25±2.6431.25_{\pm 2.64} 53.85±2.8853.85_{\pm 2.88} 64.42±11.0964.42_{\pm 11.09} 72.36±11.82
TurnOnSwitch 0.00±0.000.00_{\pm 0.00} 29.81±2.7429.81_{\pm 2.74} 55.29±2.6755.29_{\pm 2.67} 58.41±6.5558.41_{\pm 6.55} 65.38±6.81
PickStick OpenDoor 0.00±0.000.00_{\pm 0.00} 83.17±1.4483.17_{\pm 1.44} 71.15±2.1771.15_{\pm 2.17} 78.37±5.6378.37_{\pm 5.63} 86.30±5.78
PushButton 21.15±3.2721.15_{\pm 3.27} 84.13±1.3184.13_{\pm 1.31} 76.44±1.9476.44_{\pm 1.94} 79.57±5.7979.57_{\pm 5.79} 86.54±4.84
OpenMicrowave 0.00±0.000.00_{\pm 0.00} 76.92±1.8676.92_{\pm 1.86} 61.54±2.8461.54_{\pm 2.84} 77.88±6.6977.88_{\pm 6.69} 87.26±7.81
TurnOnSwitch 2.88±1.442.88_{\pm 1.44} 61.54±2.3961.54_{\pm 2.39} 48.56±3.2248.56_{\pm 3.22} 66.59±6.1766.59_{\pm 6.17} 74.52±6.10
Mean 3.09±0.723.09_{\pm 0.72} 45.61±1.9445.61_{\pm 1.94} 61.60±2.3161.60_{\pm 2.31} 69.23±1.6069.23_{\pm 1.60} 77.43±2.64

4.1 Experimental Setup

Tests use a floating Shadow Hand in Isaac Lab. Each of eight base skills is learned from 50 human demonstrations; each retention skill supplies 4,096 successful held states for composition training. Each task pair has 32 rollouts per seed across eight seeds. Success requires retaining Task A’s object throughout and completing Task B.

4.2 Comparison with Baselines

Mean success is 3.09% for frozen-grasp chaining, 45.61% for decomposed action spaces, 61.60% for a full-action residual, 69.23% for Ours-ZS, and 77.43% for full DexCompose. The improvement over the strongest external baseline is 15.83 percentage points. Ours-ZS already includes a learned retention stabilizer, so its name does not imply an entirely untrained composition system.

Original figure
Figure 3: Base-policy preservation analysis. We measure A-side and B-side preservation ratios to evaluate whether composite policies retain the capabilities of the original base policies. Frozen-Grasp preserves part of the Task-B behavior but fails to maintain Task A, while Ours achieves high preservation on both sides.

4.3 Ablation and Diagnostic Analysis

Removing the retention residual reduces average success to 9.13%; removing finger allocation yields 54.52%, and removing action masking 59.13%. Removing the downstream residual or transition stage gives 69.23% and 69.02%. These comparisons identify retention stabilization as particularly important.

Failures depend on the downstream interaction: transitions threaten stability, switch contact can disturb the grasp, and microwave actuation can fail even when retention succeeds. A four-pair mask-selection comparison favors the LLM selector, 73.0% versus 66.5%, by considering useful free fingers rather than retention alone.

Original figure
Figure 4: Failure-mode breakdown across task combinations.
Table 2: Component ablation on 16 composite tasks. Composite success rate (%), reported as mean ±\pm standard deviation over eight seeds. Ours-ZS denotes the zero-shot DexCompose variant without the Task-B Residual. The remaining columns remove one component from the full DexCompose method: −-Finger removes Finger Allocation, −-Task-A removes the Task-A Residual, −-Mask removes Action Masking, and −-Trans. removes the Transition Stage.
Composite Success Rate (%)
First Task Second Task Ours Ours-ZS −-Finger −-Task-A −-Mask −-Trans.
GraspBall OpenDoor 82.69±8.33 73.56±6.9473.56_{\pm 6.94} 73.80±7.1673.80_{\pm 7.16} 3.85±3.863.85_{\pm 3.86} 67.07±8.2367.07_{\pm 8.23} 75.00±7.9775.00_{\pm 7.97}
PushButton 85.58±7.50 76.92±4.6976.92_{\pm 4.69} 70.43±9.4370.43_{\pm 9.43} 3.37±3.713.37_{\pm 3.71} 68.75±8.3768.75_{\pm 8.37} 79.33±7.6179.33_{\pm 7.61}
OpenMicrowave 73.08±9.25 64.66±7.7064.66_{\pm 7.70} 43.51±13.2943.51_{\pm 13.29} 0.96±1.500.96_{\pm 1.50} 56.73±10.6656.73_{\pm 10.66} 64.90±9.7364.90_{\pm 9.73}
TurnOnSwitch 71.88±8.93 63.94±8.3363.94_{\pm 8.33} 54.33±8.8054.33_{\pm 8.80} 0.96±1.970.96_{\pm 1.97} 53.37±7.4953.37_{\pm 7.49} 63.22±7.5663.22_{\pm 7.56}
PourMug OpenDoor 75.72±8.08 66.59±8.7866.59_{\pm 8.78} 62.02±8.6462.02_{\pm 8.64} 15.38±5.4815.38_{\pm 5.48} 56.97±6.6556.97_{\pm 6.65} 64.90±7.8864.90_{\pm 7.88}
PushButton 85.58±7.28 78.61±6.6178.61_{\pm 6.61} 62.74±8.7062.74_{\pm 8.70} 17.79±6.5517.79_{\pm 6.55} 67.55±6.8267.55_{\pm 6.82} 76.68±6.0776.68_{\pm 6.07}
OpenMicrowave 62.74±8.21 52.40±7.6752.40_{\pm 7.67} 37.02±6.8537.02_{\pm 6.85} 13.46±5.6213.46_{\pm 5.62} 45.43±6.5845.43_{\pm 6.58} 55.05±7.7155.05_{\pm 7.71}
TurnOnSwitch 64.90±9.89 55.29±7.4855.29_{\pm 7.48} 40.62±9.2940.62_{\pm 9.29} 12.02±5.5412.02_{\pm 5.54} 44.23±7.0944.23_{\pm 7.09} 55.77±8.5455.77_{\pm 8.54}
PickCan OpenDoor 79.09±6.04 72.84±6.6872.84_{\pm 6.68} 57.21±11.3057.21_{\pm 11.30} 14.66±5.3214.66_{\pm 5.32} 60.82±8.4360.82_{\pm 8.43} 70.43±6.8270.43_{\pm 6.82}
PushButton 85.34±6.92 77.64±5.0877.64_{\pm 5.08} 56.49±10.5656.49_{\pm 10.56} 16.59±6.5516.59_{\pm 6.55} 62.26±10.4862.26_{\pm 10.48} 77.64±7.2077.64_{\pm 7.20}
OpenMicrowave 72.36±11.82 64.42±11.0964.42_{\pm 11.09} 30.29±10.1630.29_{\pm 10.16} 11.78±5.8711.78_{\pm 5.87} 51.68±10.4951.68_{\pm 10.49} 62.02±9.9562.02_{\pm 9.95}
TurnOnSwitch 65.38±6.81 58.41±6.5558.41_{\pm 6.55} 32.21±6.5532.21_{\pm 6.55} 11.54±4.8411.54_{\pm 4.84} 45.43±7.7245.43_{\pm 7.72} 57.93±6.2057.93_{\pm 6.20}
PickStick OpenDoor 86.30±5.78 78.37±5.6378.37_{\pm 5.63} 73.32±8.5273.32_{\pm 8.52} 5.53±5.135.53_{\pm 5.13} 71.88±6.3871.88_{\pm 6.38} 77.40±4.4577.40_{\pm 4.45}
PushButton 86.54±4.84 79.57±5.7979.57_{\pm 5.79} 70.91±9.2370.91_{\pm 9.23} 6.25±4.236.25_{\pm 4.23} 68.27±7.9568.27_{\pm 7.95} 78.12±4.9478.12_{\pm 4.94}
OpenMicrowave 87.26±7.81 77.88±6.6977.88_{\pm 6.69} 54.33±8.6154.33_{\pm 8.61} 3.37±2.983.37_{\pm 2.98} 68.27±6.7368.27_{\pm 6.73} 77.64±8.4577.64_{\pm 8.45}
TurnOnSwitch 74.52±6.10 66.59±6.1766.59_{\pm 6.17} 53.12±10.3653.12_{\pm 10.36} 2.88±4.132.88_{\pm 4.13} 57.45±7.7157.45_{\pm 7.71} 68.27±7.0968.27_{\pm 7.09}
Average – 77.43±2.64 69.23±1.6069.23_{\pm 1.60} 54.52±2.0954.52_{\pm 2.09} 9.13±1.299.13_{\pm 1.29} 59.13±1.7559.13_{\pm 1.75} 69.02±2.0769.02_{\pm 2.07}
Table 3: Heuristic vs. LLM agent mask selection.
Task Pair Heur. LLM
GraspBall+OpenDoor 72.1 82.7
GraspBall+Switch 63.5 71.9
PickCan+Micro. 68.3 72.4
PourMug+Switch 62.0 64.9
Mean 66.5 73.0

5 Conclusion

Explicit action ownership and learned stabilization enable two-skill compositions while reusing frozen base policies. The evidence covers these simulated task pairs. Extending to physical hands or longer sequences with several interacting skills remains untested here.

Appendix

The appendices specify base-policy training, residual learning, release tests, comparison implementations, and the mask-selection prompt.

Appendix A Base Policy Training and Evaluation Protocol

Base policies use 32–37 low-dimensional proprioceptive and geometric state features, rather than raw vision. Actions contain six wrist coordinates and 22 finger-joint targets. A FiLM-conditioned temporal UNet predicts 32 actions from two observations; execution horizons vary by task.

Training uses successful demonstration windows, a 100-step diffusion objective, AdamW at 10⁻⁴, batch size 256, a 2,000-step warmup, cosine decay, EMA, and gradient clipping. Table 4 gives standalone success from 82.21% to 100%. Composite evaluation additionally requires continuous retention while the interaction reaches its geometric success threshold.

Equations
at=(pt,rt,qt),a_{t}=(p_{t},r_{t},q_{t}),
Linear→Mish→Linear,\text{Linear}\rightarrow\text{Mish}\rightarrow\text{Linear},
o​b​s∈ℝ2×do​b​s,a​c​t​i​o​n∈ℝ32×28.obs\in\mathbb{R}^{2\times d_{obs}},\qquad action\in\mathbb{R}^{32\times 28}.
Table 4: Standalone success rate (%) of pretrained base policies. Results are evaluated using the corresponding execute horizon for each task.
Task Execute Horizon Success Rate (%)
GraspBall 8 100.0
PickStick 16 100.0
PickCan 16 100.0
PourMug 16 100.0
OpenDoor 24 97.59
OpenMicrowave 24 89.90
PushButton 16 100.0
TurnOnSwitch 30 82.21
Equation
Scomp=SA∧SB.S_{\mathrm{comp}}=S_{A}\land S_{B}.

Appendix B Learning Procedure and Implementation Details

The implementation uses one RTX 4090. Both residuals are learned with PPO; the allocation stage uses release-test statistics and an LLM rather than gradient training.

B.1 PPO Training: Task A Residual Stabilizer

The retention residual observes preserved-finger motion, object state in the palm frame, contacts, torque features, executed base action, and rollout phase. Its bounded corrections minimize object–palm distance plus a 0.001 residual penalty. A 256–256–128 ELU backbone trains for 1,000 PPO iterations in 1,024 environments, totaling 24,576,000 steps and about 20 minutes.

Equation 13
RtA=−‖𝐩obj−𝐩palm‖2−α​‖ΔtA‖22,R^{A}_{t}=-\|\mathbf{p}_{\mathrm{obj}}-\mathbf{p}_{\mathrm{palm}}\|_{2}-\alpha\|\Delta^{A}_{t}\|^{2}_{2}, (13)

B.2 PPO Training: Task B Residual

The downstream residual additionally receives the frozen Task-B action and selected mask. Reward combines task progress, retention distance with weight one, and residual penalty 0.001. Optional joint updates adapt the Task-A residual at a reduced learning rate while base policies remain frozen.

Training uses eight environments, 200-step nominal rollouts, and 1,000 iterations: 1.6 million nominal step slots, with filtering reducing actual samples per update to roughly 408–750. Reported runtime is about four hours.

Equation 14
RtB=RtaskB​(st,at)−β​‖𝐩obj−𝐩palm‖2−α​‖δtB‖22,R^{B}_{t}=R_{\mathrm{taskB}}(s_{t},a_{t})-\beta\|\mathbf{p}_{\mathrm{obj}}-\mathbf{p}_{\mathrm{palm}}\|_{2}-\alpha\|\delta^{B}_{t}\|^{2}_{2}, (14)

B.3 Training Curriculum

First collect held states; then train one retention stabilizer per Task-A skill under randomized disturbances. For each task pair, compute release diagnostics, choose a fixed mask, and train a downstream residual. Thus reuse avoids retraining full base policies, but additional pair-specific learning remains.

B.4 Release-Test Protocol

Each candidate mask receives 100 tests from randomly restored held states. Released fingers open over 100 simulation steps. Appendix retention checks use a five-centimeter palm-distance threshold and three-centimeter drop-height threshold at the end; clean release requires contact forces below 0.1 newton during the final twenty steps. Composite evaluation separately requires retention throughout execution.

Equation 15
λt=tTtest,\lambda_{t}=\frac{t}{T_{\mathrm{test}}}, (15)

B.5 Baseline Implementation Details

Frozen grasp holds finger targets fixed without residual correction. The decomposed baseline retrains base policies with fixed family-specific masks and cross-subspace action penalties. The residual baseline uses a single unrestricted correction trained with SAC; it differs in learning algorithm as well as ownership structure. Ours-ZS uses the learned retention controller and selected mask but no learned downstream correction.

Equation 16
at=MA⊙aAref+MB⊙πB​(st)⏟nominal+δt,a_{t}=\underbrace{M_{A}\odot a^{\mathrm{ref}}_{A}+M_{B}\odot\pi_{B}(s_{t})}_{\text{nominal}}+\delta_{t}, (16)
Comparison of heuristic and language-model finger allocation for holding a ball while opening a door.
Figure 5: Comparison between heuristic and LLM-based mask selection on the GraspBall+OpenDoor task. The heuristic chooses a thumb–index grasp because it maximizes immediate object retention, but the occupied index finger prevents effective door interaction in the subsequent stage. In contrast, the LLM selects a thumb–ring–little grasp that preserves the index finger for future manipulation.

B.6 Hyperparameters and Compute Resources

PPO uses learning rate 3×10⁻⁴, clipping 0.2, discount 0.99, GAE parameter 0.95, five epochs, four minibatches, and gradient norm limit one. Optional Task-A joint tuning uses 10⁻⁴. Both actors start with zero-mean corrections and log standard deviation −1.5. The separate training budgets and release thresholds are retained in their preceding sections.

B.7 LLM-Based Finger Mask Selection

The authors use GPT-5.4 with temperature zero to choose a mask from task descriptions, candidate masks, retention/clean-release rates, and free-finger counts. The original prompt is retained below. No model fine-tuning is used at this stage; low-temperature decoding is a configuration choice rather than a general reproducibility guarantee.

The heuristic maximizes retention subject to releasing at least two fingers. In the illustrated ball-and-door example, freeing the index finger helps downstream contact even though its selected grasp has lower immediate retention. This explains the measured benefit on selected pairs without establishing that an LLM is necessary for every allocation.

“You are selecting a finger-retention mask for a dual-task robotic manipulation problem. The robot hand has five fingers and must continue stabilizing the object from Task A while freeing sufficient fingers to execute Task B. Each candidate mask specifies which fingers remain in contact with the object and which fingers are released for downstream interaction.

A good mask should satisfy three objectives simultaneously: (1) maintain stable object retention, (2) allow clean finger release without disturbing the grasp, and (3) preserve dexterous and anatomically appropriate fingers for Task B.
When evaluating candidates, consider not only the numerical metrics but also the semantic suitability of the finger assignment. For example, thumb-index opposition is generally preferable for stabilizing slender cylindrical objects, while releasing the thumb may significantly reduce grasp stability even if the clean-release score is high. Avoid overly conservative masks that release too few fingers and overly aggressive masks that sacrifice object stability.
Select the single mask that provides the best overall trade-off between grasp stability and downstream manipulation capability.”
Equation 17
mheur=arg⁡maxm∈ℳ​Pret​(m)s.t.5−‖m‖0≥2.m_{\mathrm{heur}}=\arg\max_{m\in\mathcal{M}}P_{\mathrm{ret}}(m)\quad\text{s.t.}\quad 5-\|m\|_{0}\geq 2. (17)
Concept demonstration

Open image ↗