acceptodds
Under review as a conference paper at ICLR 2027

PanTact: Teaching Whole-Body Dexterity with a Wearable Contact Interface

Abstract

Humanoid robots have many degrees of freedom and contact surfaces, yet most loco-manipulation research restricts manipulation to the hands and avoids other body contacts with the environment. To extend manipulation beyond the hands, we present PanTact, a system that learns humanoid whole-body dexterity, the capability to manipulate with any suitable body surface, directly from robot-free human demonstrations. Our key idea is to use contact as manipulation-centric information that transfers across embodiments. Contact locations and timing specify the interactions needed to accomplish a task, while the robot chooses how to coordinate its body to realize them within its own physical constraints. Specifically, PanTact captures, learns, and executes whole-body contact via: 1) a wearable, robot-free data collection interface that augments egocentric vision and bimanual UMI grippers with whole-body modular contact patches to record body contact locations and states; 2) a visuomotor policy that mitigates the egocentric observation gap and predicts whole-body contact goals; and 3) contact-aware whole-body control formulations that realize active contact goals using suitable robot surfaces. We deploy PanTact on a Unitree G1 humanoid without additional body contact sensors, demonstrating whole-body dexterity involving body support, pushing, sweeping, and load-sharing.

Primary area:
applications to robotics, autonomy, planning
Keywords:
humanoid loco-manipulation, whole-body manipulation, scalable demonstration interface, tactile sensing, imitation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_6fm4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents PanTact, a system for teaching a humanoid to manipulate with any suitable body surface (forearms, knees, legs), learned from robot-free human demonstrations. Its central idea is to treat contact as a cross-embodiment interface: the human demonstration specifies where and when contact should occur, and the robot decides how to realize it with its own body. The system has three parts. 1. A wearable data-collection rig that adds modular, magnetically tiled FSR contact patches with Vive trackers on the forearms and knees to an egocentric, bimanual UMI setup. 2. A diffusion policy that takes a body-masked egocentric pointmap and wrist images as input and predicts end-effector (EEF) trajectories together with per-sensor contact goals (binary state and 3D position). 3. Three contact-aware whole-body controllers that realize active contact goals on the closest robot surface, using a soft-min distance and a virtual-target offset to generate pressing force: upper-body differential IK, whole-body IK tracked by SONIC, and an RL-trained sparse contact tracker. On a Unitree G1 without body sensors, PanTact succeeds on Grasp-and-Embrace (85% vs. 0% for Humanoid UMI), Sweep-and-Stow (90.6% vs. 0% end-to-end for Humanoid UMI++), and Carry-and-Push (90% vs. 40%). A simulation replay study compares the controllers with and without contact goals.

Critical strengths
  1. The paper tackles a well-motivated problem that exploits what makes humanoids distinctive. Most humanoid loco-manipulation work reduces the robot to two grippers on a mobile base. This paper targets manipulation with the forearms, knees, and legs, a capability that only a human-like morphology makes possible and that the hands alone cannot provide. Treating contact location and timing as the quantity that transfers, while leaving body configuration to the robot, is a clean and well-argued way to bridge the human–robot morphology gap. All three tasks genuinely require non-hand contact (embracing a bulky object, pushing a cabinet closed with a knee or foot, load-sharing while pushing with a leg).

  2. The system is complete and tested on a real robot, with large gains over pose-based interfaces. The pipeline covers wearable hardware, policy, and controller, and is deployed on a commercial humanoid without adding body sensors. Pose-only baselines fail in characteristic ways: they never establish sufficient arm contact, or they drive the arms into joint limits. The gains are large (0 → 85%, 0 → 90.6%, 40 → 90%). Together with the planned release of the modular patch design, this is a useful contribution to the community.

Critical weaknesses
  1. Execution quality is not convincing, and how to realize contact on the robot remains unresolved. The supplementary videos show visibly jittery, non-smooth motion, yet the paper reports only binary success rates and median tracking errors, with no measure of smoothness, jerk, or contact stability. Several design choices could plausibly cause this behavior, and none are analyzed: contact goals are binarized FSR readings, so the controller's contact objective switches on and off abruptly; pressing force is produced open-loop by offsetting a position target into the object (Eq. 4), with no force feedback on the robot; and a 10 Hz diffusion policy is stitched asynchronously into a 50 Hz tracker. The tracking plots (Figs. 4e and 6e) also show noisy EEF and contact traces, and the sparse tracker is reported to have "tracking dead zones." The paper also proposes three controllers without settling on one. Each task uses a different one (Upper-Body IK for Grasp-and-Embrace, WBIK+SONIC for the other two), and Fig. 7 shows a trade-off in which none dominates. Since contact-aware control is presented as a core contribution, the reader should learn which formulation to use and why. As it stands, the robot-side execution looks like a set of options picked per task rather than a solved component, and the jitter makes it unclear how robust the whole-body behaviors are beyond the reported trials.

  2. The evaluation is too narrow to isolate the value of contact prediction. The evaluation uses three tasks with 10–20 trials each, in a single lab setting. The baselines also differ from PanTact in more than the use of contact information. For Grasp-and-Embrace, the only baseline (Humanoid UMI) has no information about arm posture at all, so its 0% mainly shows that the arms need some target. The paper omits the most informative ablation: track the forearm patch poses, which are already recorded by the Vive trackers on the patches, continuously and without contact states. Humanoid UMI++ does this only for the knees. There are also no ablations of the policy itself (body masking of the egocentric view, predicting contact goals vs. deriving them from predicted poses, history length), and the recorded contact intensity is never used. The paper therefore does not show which part of "contact as interface" drives the gains. On the learning side, the policy is a standard diffusion policy following HoMMI with an extended action space, so the contribution is mostly at the level of hardware and systems. The evidence should be correspondingly thorough, but it is not.

Questions for authors

Can you quantify motion quality, for example joint/EEF jerk, contact-state switching frequency, or contact-force variance during active contact, for each controller and for the baselines? What causes the jitter visible in the videos: flickering contact goals, the virtual-target offset, the asynchronous chunk scheduling, or the trackers themselves?

1 comment

Thank you for the constructive feedback and for recognizing the strong motivation for whole-body dexterity, the contact interface for human-to-robot transfer, and our complete real-robot system. We address your concerns about execution quality, controller selection, and the contributions of contact prediction below.

Motion quality

The front-page videos play at 4× speed, which makes small adjustments appear faster and more abrupt. Several non-smooth motions reflect the robot responding to feedback during the task and attempting retrying behaviors during complex, long-horizon tasks. In Sweep-and-Stow, the robot adjusts its arm position relative to the objects as it sweeps them into the bin. In Carry-and-Push, it can press its knee again when the first attempt does not move the floor box sufficiently. The policy also replans as objects become occluded or come back into view. The task videos play at 2x speed, and the additional videos play at 1× speed to show motion quality more clearly.

We quantify motion quality from real-world Grasp-and-Embrace rollouts.

MetricUpper-Body IKSparse Contact Tracker
Arm-joint jerk RMS [rad/s³]71.945.9
EEF jerk [m/s³]20.320.7
Upper-body velocity jitter above 4 Hz [rad/s]0.0390.027
Contact-command toggles per family [1/s]0.5–1.11.1–1.6

The Sparse Contact Tracker achieves 36% lower arm-joint jerk, while the EEF jerk is similar. Contact transitions contribute to arm reconfiguration as the controller accommodates contact goals alongside wrist targets. The Sparse Contact Tracker smooths target changes through action filtering, softer gains, and action-rate regularization.

Contact stability

We run additional experiments comparing different pressing parameters and applying disturbance forces to objects. The robot maintains stable contacts under the applied disturbances.

Controller selection

We use Upper-Body IK for Grasp-and-Embrace, where the robot can keep its feet fixed, and WBIK+SONIC for the other two tasks, which requires coordinated lower-body manipulation, locomotion, and balance.

We recommend upper-Body IK for fixed-leg upper-body tasks, where it achieves the lowest EEF error. For whole-body tasks, controller selection depends on the task requirements. On Carry-and-Push, WBIK+SONIC achieves 90% success versus 60% for the Sparse Contact Tracker. Its lower EEF error helps maintain the bimanual support configuration during carrying is crucial to this task’s success, while leg contacts tolerate larger tracking errors.

Contact interface ablations

We ran two additional baselines. The first is a continuous tracking baseline that tracks the same contact points continuously using the any-surface objective, without gating by contact states. It achieves 4/10 success (40%), compared with 9/10 (90%) for ours. The failure cases videos show box slips and walking in the wrong direction. The second, a contact-gated fixed-link baseline, retains the same contact positions and states but assigns each goal to a fixed body part. It achieves 5/10 (50%), with box slips and missed leg contacts. These results show that both contact gating and adaptable surface assignment contribute to task success.

MethodContact gatingSurface assignmentSuccess
Humanoid UMI++ (original)NoFixed link4/10 (40%)
Continuous tracking (new)NoAny surface4/10 (40%)
Contact-gated fixed-link (new)YesFixed link5/10 (50%)
Ours (original)YesAny surface9/10 (90%)

Policy ablations

We additionally train policy variants on the same Grasp-and-Embrace dataset and evaluate them on the same 5% held-out validation split. The table reports changes in position error relative to the original policy. Positive values indicate higher error.

Policy variantLeft wristRight wristArm contacts (L / R)
One history frame+3.85 mm+1.36 mm+3.5 / +2 mm
Contact positions derived from poses00+0.22 / +0.22 mm

Using one history frame increases prediction error. Deriving forearm-contact positions from predicted wrist poses gives comparable validation accuracy to predicting them directly.

Full task completion

For Sweep-and-Stow, we originally used partial credits because the baseline would be a rather 0% success and less informative to analyze failure modes. We rescore the original rollouts and ours fully completes 13/16 rollouts (81.3%), compared with 0% for Humanoid UMI++.

MethodFully sweptFully closedFull end-to-end completion
Humanoid UMI++1/160/160/16
Ours16/1613/1613/16 (81.3%)
Reviewer_YC57OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

PanTact learns whole-body manipulation from robot-free human demonstrations. Wearable contact patches record where and when body contact occurs, a visuomotor policy predicts contact goals alongside hand motions, and a whole-body controller realizes them on a G1 humanoid.

Critical strengths

s1. Contact goals offer a useful way to teach interactions that hand trajectories alone leave unspecified.

s2. The system connects wearable sensing, policy learning, and control, with real-robot demonstrations of arm support, sweeping, and leg pushing.

s3. Deployment requires no additional body-contact sensors on the robot, which makes the approach more practical.

Critical weaknesses

w1. The baselines show that contact information helps, but do not isolate the benefit of letting the controller choose the body surface. A baseline using the same contact goals with fixed body assignments would be more informative.

w2. Real-robot evaluation uses only 10–20 trials per task. The Sweep-and-Stow score also includes partial completion, making it harder to judge how often the full task succeeds.

w3. The sparse tracker achieves smaller contact-position errors but performs worse on Carry-and-Push. The current analysis does not fully explain which tracking errors matter most for task success.

w4. The method primarily optimizes the geometric distance between target contact points and robot body surfaces, without explicitly modeling contact forces, friction, or multi-contact load distribution. Although virtual contact targets are introduced to encourage pressing forces, the lack of closed-loop force feedback cannot guarantee stable and effective physical interactions. Consequently, the method mainly captures contact geometry rather than full contact dynamics, potentially limiting its robustness in contact-rich manipulation tasks.

Questions for authors

q1. Could you compare against a controller that uses the same contact positions and activation flags but assigns each goal to a fixed body part?

q2. How much does tracking contact points only when active help? Could you compare it with continuously tracking the same points?

q3. Could you report full-completion counts separately from partial scores for Sweep-and-Stow, and add trials for the smaller evaluations?

q4. Although virtual contact targets are introduced to encourage pressing forces, how does the method ensure stable and effective physical interactions without closed-loop force feedback? Could the authors provide additional evaluations on contact stability and robustness under varying object masses and friction conditions?

1 comment

Thank you for the constructive feedback and for recognizing the value of contact goals, our integrated sensing, learning, and control system, and deployment without additional body-contact sensors. We address your questions about body assignment, contact gating, full task completion, controller tradeoffs, and contact stability below.

w1/q1. Same contact goals with fixed body assignments

We ran an additional contact-gated fixed-link baseline that uses the same contact positions and activation flags but assigns each goal to a fixed body part. It achieves 5/10 success (50%), compared with 9/10 (90%) for ours. The robot sometimes drops the carried box or fails to reach the floor box with its leg. Simultaneously tracking wrist poses and fixed arm keypoints can overconstrain the robot and compromise arm support. Fixed body assignments also restrict how the robot reaches each contact goal, leaving the robot surface short of the intended contact when human and robot body proportions differ. The fixed-link failure cases videos illustrate these failures.

q2. Continuously tracking contact points

We also ran a new continuous tracking baseline that tracks the same contact points continuously using the same any-surface objective, without gating by contact states. It achieves 4/10 success (40%), compared with 9/10 (90%) for ours. Without contact states, the controller maintains contact target tracking constraints during non-contact free-space motion. These constraints can interfere with hand tracking and whole-body coordination. The continuous-tracking failure cases videos show box slips and walking in the wrong direction.

MethodContact gatingSurface assignmentSuccess
Humanoid UMI++ (original)NoFixed link4/10 (40%)
Continuous tracking (new)NoAny surface4/10 (40%)
Contact-gated fixed-link (new)YesFixed link5/10 (50%)
Ours (original)YesAny surface9/10 (90%)

w2/q3. Full-completion counts

For Sweep-and-Stow, we originally used partial credits because the baseline would be a rather 0% success rate and less informative to analyze failure modes. We rescore the original 16 rollouts per method using a criterion that requires sweeping all objects into the bin and fully closing the cabinet Ours fully sweeps all objects in 16/16 rollouts, fully closes the cabinet in 13/16, and fully completes the task in 13/16 (81.3%). Humanoid UMI++ achieves 1/16 full sweeps and 0/16 full completions.

MethodFully sweptFully closedFull end-to-end completion
Humanoid UMI++1/160/160/16
Ours16/1613/1613/16 (81.3%)

w3. Which tracking errors matter for task success?

The controller formulations have different trade-offs, as discussed in the main paper. The relationship between tracking error, contact distance, and task success depends on the task requirements. Tasks that require higher end-effector tracking accuracy favor Upper-body IK or WBIK+SONIC, while tasks that prioritize body-contact accuracy favor the Sparse Contact Tracker. Carry-and-Push requires the robot to maintain a bimanual support configuration while walking and pushing. WBIK+SONIC achieves 90% success, compared with 60% for the Sparse Contact Tracker, whose larger EEF tracking error more often leads to dropping the carried box. However, the Sparse Contact Tracker pushes the floor box faster and farther with its legs, reflected by its smaller contact distance.

w4/q4. Contact stability and robustness

The policy runs in closed loop and uses proprioception and visual feedback to sustain contact and adjust the interaction. The low-level controller tracks a virtual target displaced by δ beyond the intended contact location along the pressing direction. When the object resists further motion, the remaining tracking error produces a pressing effort. The visuomotor policy observes the task outcome and updates the hand trajectories and contact goals. For example, it can command another knee push when the first attempt leaves the floor box short of the goal.

To further evaluate contact stability, we apply external disturbances by pushing directly on the box while the robot holds it. The robot maintains arm support and retains the box throughout the shown external-push test videos. We also compare offsets of 0, 0.025, 0.05, and 0.075 m on Carry-and-Push. Smaller offsets can produce insufficient floor-box motion, while 0.075 m produces a more aggressive push. The offset comparison video shows these behaviors.

Reviewer_3iUsOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PanTact learns humanoid manipulation that uses forearms and lower-body surfaces, rather than only grippers. A robot-free demonstration interface combines egocentric/wrist cameras, tracked UMI grippers, and modular force-sensitive contact patches. A diffusion policy predicts end-effector commands and contact activation/location goals from body-masked visual observations. Three controller variants realize active goals using suitable robot surfaces: upper-body IK, whole-body IK with SONIC execution, and a learned sparse contact tracker. Robot deployment requires no added body-contact sensors; pressing effort is encouraged through virtual target offsets. On a G1, the paper studies Grasp-and-Embrace, Sweep-and-Stow, and Carry-and-Push with 100, 150, and 110 demonstrations, respectively. It reports 85% binary success on Grasp-and-Embrace, a 90.63% graded end-to-end score on Sweep-and-Stow, and 90% binary success on Carry-and-Push, alongside pose-only baselines and simulated contact-tracking comparisons.

Critical strengths
  1. This paper proposes a consequential task-level interface from human sensing to robot execution. The contribution is not simply a wearable sensor: measured contact timing and locations supervise policy outputs that are explicitly consumed by the controller. Activating only task-relevant contacts and allowing different robot surfaces to realize them addresses a real human-to-robot morphology mismatch.

  2. This paper presents compelling physical demonstrations with task-relevant comparisons. The experiments go beyond free-space motion tracking to support, sweeping, load sharing, and leg-assisted pushing. Grasp-and-Embrace reaches 85% success versus 0% for Humanoid UMI, while Carry-and-Push reaches 90% versus 40% for Humanoid UMI++ in the reported trials

Critical weaknesses
  1. The physical realization of contact force is not fully specified or measured. Equation 4 requires a pressing direction and nominal force/stiffness/offset parameters, but the policy output in Equation 2 contains only binary activation and 3D position. The manuscript does not explain sufficiently how the pressing direction is recovered online, which force parameters are used, or how sensitive execution is to those choices.

  2. Pose-only baselines differ from PanTact in contact information and how tracking constraints are activated/assigned. Figure 7 usefully controls contact-goal presence at the controller level, but does not distinguish the benefit of learned contact states, geometric body masking, and any-surface realization in closed-loop hardware tasks. In particular, a contact-gated fixed-link baseline would test whether continuous knee-pose tracking is the main limitation of UMI++. The 10-20 rollouts per task and graded cabinet-closure metric further limit precision about generalization and fully completed task frequency.

Questions for authors

See weaknesses.

1 comment

Thank you for the constructive feedback and for recognizing our task-level contact interface and compelling physical demonstrations. We address your concerns about contact-force realization, the benefits of contact-gating and surface assignment, and full task completion below.

Pressing direction, force parameters, and sensitivity

The pressing direction is calculated from the normalized vector from the closest surface to contact targets in the current robot configuration toward the contact target.

We encourage pressing by placing the virtual target a distance δ beyond the intended contact location along the pressing direction. Equation 4 expresses this displacement as δ = min(F_des/k_virt, δ_max). We tune the displacement directly in deployment and use δ = 0.05 m for the original Carry-and-Push trials. The low-level controller tracks the offset target, producing a pressing effort while the object constrains the robot surface.

We ran additional Carry-and-Push experiments with offsets of δ = 0 m, δ = 0.025 m, and δ = 0.075 m, using the original δ = 0.05 m setting as the reference. At zero offset, the controller aims for the intended contact location without an additional pressing displacement. Smaller offsets can leave the robot unable to push the floor box sufficiently. Increasing δ to 0.075 m places the virtual target farther into the object and produces a more aggressive forward push.

We also test whether the robot maintains stable contacts under external disturbance forces. With the Carry-and-Push controller running, a person pushes directly on the box between the robot's arms. The robot maintains arm support and keeps holding the box throughout the shown perturbation, as demonstrated in the hold-under-push video.

The benefits of learned contact states and any-surface realization

We ran an additional contact-gated fixed-link baseline that uses the same contact positions and activation flags but assigns each goal to a fixed body part. It achieves 5/10 success (50%), compared with 9/10 (90%) for ours. The robot sometimes drops the carried box or fails to reach the floor box with its leg. Tracking wrist poses and fixed arm keypoints together over constrains the robot and compromises arm support, while the difference between human and robot body proportions and tracking error can leave the robot surface short of the intended contact. The fixed-link failure cases videos show both outcomes.

We also ran a continuous tracking baseline that continuously tracks the same contact points using any-surface objective, without gating with contact states. It achieves 4/10 success (40%), compared with 9/10 (90%) for ours. Tracking points during non-contact free-space motion adds constraints that can interfere with hand tracking and whole-body coordination. The continuous-tracking failure cases videos show box slips and walking in the wrong direction.

These two additional baselines together show that the designs of contact gating and adaptable body assignment are both important.

MethodContact gatingSurface assignmentSuccess
Humanoid UMI++ (original)NoFixed link4/10 (40%)
Continuous tracking (new)NoAny surface4/10 (40%)
Contact-gated fixed-link (new)YesFixed link5/10 (50%)
Ours (original)YesAny surface9/10 (90%)

Full completion metric on sweep-and-stow task

We originally gave partial credits because the baseline would be a rather 0% success rate and less informative to analyze failure modes. We rescore the original 16 Sweep-and-Stow rollouts per method using a criterion that requires sweeping all objects into the bin and fully closing the cabinet. Ours achieves 13/16 full completions (81.3%), compared with 0/16 for Humanoid UMI++. Ours sweeps all objects in 16/16 rollouts and fully closes the cabinet in 13/16.

MethodFully sweptFully closedFull end-to-end completion
Humanoid UMI++1/160/160/16
Ours16/1613/1613/16 (81.3%)

Comments on the paper

1 comment

We thank the reviewers for their constructive feedback and for recognizing the value of whole-body dexterity, contact as a cross-embodiment interface, and our complete system demonstrated end-to-end from data collection to policy deployment on a real humanoid. We address the shared concerns with two additional ablation studies on the contact interface between contact-predictive policy and contact-aware controller, full task-completion counts for Sweep-and-Stow, additional evaluations of sensitivity to pressing parameters and object loads, and policy and motion-quality analyses. We have added supporting videos to the project’s submission website https://pantacticlr.github.io.

Contact gating and surface assignment ablations

We ran two additional baselines on the Carry-and-Push task using the same demonstrations and evaluation configurations, with 10 rollouts each. The Continuous tracking baseline retains the same contact points and any-surface objective but removes contact gating and lets the whole-body controller continuously track all the contact points predicted by the policy. Contact-gated fixed-link retains the contact goals and activation flags but fixes the body assignment to each contact target.

MethodContact gatingSurface assignmentSuccess
Humanoid UMI++ (original)NoFixed link4/10 (40%)
Continuous tracking (new)NoAny surface4/10 (40%)
Contact-gated fixed-link (new)YesFixed link5/10 (50%)
Ours (original)YesAny surface9/10 (90%)

Removing gating reduces success from 90% to 40%; fixing the body assignment reduces it to 50%, therefore both components contribute to the successful cross-embodiment contact transfer. The contact-gated fixed-link baseline sometimes drops the carried box or fails to reach the floor box with its leg. The controller must track wrist poses and fixed arm keypoints at the same time, which can prevent the arms from maintaining sufficient support. Fixed body assignments also restrict how the robot reaches each contact goal. Differences in body proportions and tracking error can leave the robot surface short of the intended contact location. The fixed-link failure-case videos show both box slips and missed leg contacts.

The continuous tracking baseline sometimes drops the carried box or walks in the wrong direction. Without contact states, the controller tracks the same contact points even during non-contact motion. These extra constraints can interfere with hand tracking and whole-body coordination, making it harder for the robot to support the box and move in the intended direction. The continuous-tracking failure-case videos show box slips and drifted robot posture.

Full Sweep-and-Stow task completion

Rescoring the original rollouts using full completion rather than partial credit gives:

MethodFully sweptFully closedFull end-to-end completion
Humanoid UMI++1/160/160/16
Ours16/1613/1613/16 (81.3%)

The baseline never fully closes the cabinet, while ours fully completes 8/8 seen object and 5/8 unseen object rollouts, where the three failures are partial cabinet closures.

Sensitivity to pressing parameters and object loads

We compare virtual target offsets of δ = 0, 0.025, 0.05, and 0.075 m on Carry-and-Push. The original trials use δ = 0.05 m. Smaller offsets can leave the floor box insufficiently displaced, while the larger offset produces a more aggressive push, as shown in the pressing-parameter sensitivity videos. We also push the box while the robot holds it between its arms. The robot maintains its hold under the external push, as shown in the contact stability video.

Policy ablations

We additionally train policy variants on the same Grasp-and-Embrace dataset and evaluate them on the validation set not seen during training. The table reports changes in position error relative to the original policy. Positive values indicate higher error.

Policy variantLeft wristRight wristArm contacts (L / R)
One history frame instead of two+3.85 mm+1.36 mm+3.5 / +2 mm
Contact positions derived from predicted wrist poses00+0.22 / +0.22 mm

Using one history frame increases prediction error on every validation episode. Deriving forearm-contact positions from predicted wrist poses gives comparable validation accuracy to predicting them directly. The policy-ablation results are also on the project website.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.