💻 computer science

Transferring Contact, Not Just Motion: Compliant Grasping Across Dexterous Hands

This paper introduces a cross-embodiment force-position interface that calibrates heterogeneous dexterous hands to share contact-aware load descriptors alongside motion intent, enabling the transfer of compliant grasping policies across different robotic systems.

Original authors: Soofiyan Atar, Yao-Ting Huang, Michael Yip

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Soofiyan Atar, Yao-Ting Huang, Michael Yip

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to pick up a fragile, squishy cookie with a robotic hand. If the robot just moves its fingers to the "right" shape, it might crush the cookie or drop it the moment it slips. To do this well, the robot needs to feel how hard it is squeezing, not just where its fingers are.

This paper introduces a new way to teach different types of robotic hands to "feel" and "grip" objects, even if the hands look completely different from one another. Here is the breakdown using simple analogies:

1. The Problem: Different Hands, Different "Languages"

Think of robotic hands like different musical instruments. One might be a piano (many keys), another a guitar (fewer strings), and a third a drum kit.

The Old Way: Previous methods tried to teach these instruments to play the same notes (finger positions). But if the piano player presses a key too hard, the note is loud; if the guitar player presses too hard, the string breaks. The old methods didn't translate the force (how hard they are pressing), only the position (where they put their fingers).
The Result: A robot hand trained on a piano couldn't easily switch to a guitar because it didn't know how to adjust its "touch" for the new instrument.

2. The Solution: A Universal "Force Translator"

The authors created a Universal Force-Position Interface. Think of this as a translator that speaks two languages at once:

The "Where" Language (Motion): They use a shared "latent code" (a compressed digital blueprint) based on human hand shapes. This tells every robot hand where to move, regardless of its size or shape.
The "How Hard" Language (Force): This is the big innovation. They take the raw, confusing signals from each specific robot hand (like "motor effort") and translate them into physical torque (Newton-meters).
- Analogy: Imagine every robot hand has a different type of speedometer. The authors built a converter that turns every speedometer reading into "Miles Per Hour." Now, a robot with a "giant" speedometer and a robot with a "tiny" speedometer both understand exactly how fast they are going.
- They also created a "Load Descriptor," which is like a map showing exactly where on the finger the object is touching and how the weight is distributed.

3. The Brain: MARC (The "Blindfolded" Learner)

They trained a new AI policy called MARC (Mask-Aware Reactive Compliant).

The Training Trick: During training, they sometimes put a "blindfold" (visual masking) over the robot's cameras. This forces the robot to stop relying on sight and learn to use its "sense of touch" (the calibrated force data) to figure out what to do.
The Result: The robot learns that if it can't see the object clearly (because its own hand is blocking the view), it should trust the pressure sensors on its fingers to know if it's holding on tight enough or if it's about to drop something.

4. The Execution: A Hybrid Team

The system doesn't just use one brain for everything. It uses a team approach:

The Learner (MARC): Handles the tricky part: reaching out and grabbing the object. It uses the "blindfolded" training to be safe and gentle.
The Muscle (Model-Based Controller): Once the object is grabbed, a simpler, math-based controller takes over to hold it steady while the robot moves it. If the robot bumps into something, this controller automatically loosens the grip just enough to prevent damage, then tightens again.
The Handover: When giving the object to a human, the system uses the force sensors to know exactly when to let go, rather than just opening the hand on a timer.

5. The Results: One Brain, Many Hands

The team tested this on three very different robotic hands (some with 5 fingers, some with fewer, some with different joint structures).

The Magic: They trained the system on one set of hands and then tested it on a new hand configuration they had never seen before.
The Outcome: The system worked surprisingly well. By translating the "force" into a universal language, the robot could transfer its skills from a 20-joint hand to a 15-joint hand without needing to relearn everything from scratch.
Success Rate: On difficult tasks (like picking up a slippery egg or a stack of cups), the success rate jumped significantly (from about 44% to 71%) when they used this new "force translator" compared to old methods that only looked at finger positions.

Summary

In short, the paper says: "Don't just teach robots where to move; teach them how hard to squeeze, and translate that feeling into a universal language so any robot hand can understand it." This allows robots to handle delicate, slippery, or squishy objects safely, even if they are using a completely different type of hand than the one they were trained on.

Technical Summary: Transferring Contact, Not Just Motion

Problem Statement

Dexterous manipulation requires more than predicting finger motion; it necessitates the active regulation of contact forces to handle slipping, deformation, and self-occlusion. Existing cross-embodiment policies successfully unify motion intent through retargeted poses or latent action spaces, but they fail to transfer contact feedback. Force feedback remains tied to hand-specific sensing and actuation (e.g., raw effort signals or tactile proxies), making it impossible to compare loading across heterogeneous robotic hands. Consequently, policies trained on one hand cannot generalize to another when visual cues are occluded or when precise force regulation is required for fragile objects.

Methodology

1. Unified Force-Position Interface

The core contribution is a transferable interface that decouples motion intent from contact loading, expressing both in physical, comparable units.

Shared Motion Intent: Hand pose is represented in a shared MANO-pose latent ( $\psi \in \mathbb{R}^{15}$ ). This low-dimensional vector is learned once from human glove data and mapped to any specific hand's joint configuration ( $q^{(h)}$ ) via a hand-specific retargeting map ( $R_h$ ) that optimizes for fingertip position and finger shape.
Calibrated Contact Channel: Instead of using raw effort, the system performs system identification to map hand-specific effort proxies to physical joint torques ( $\tau^{(h)}$ $τ^{(h)}$ in N·m). From these torques, the system derives:
- Fingertip Forces ( $f_t$ ): Calculated via the least-squares inverse of the translational Jacobian.
- Spatial Torque Descriptors ( $\gamma_t$ ): A compact set of three scalars ( $c, \rho, \ell$ ) encoding the contact location along the finger, the distribution of load across joints, and the lateral share of the force.
  This allows the policy to observe "where the hand should move" and "how the object is loaded" in a unified coordinate system, regardless of the specific hand embodiment.

2. MARC: Mask-Aware Reactive Compliant Policy

The authors introduce MARC, a flow-matching policy based on a Diffusion Transformer (DiT) architecture.

Inputs: The policy is conditioned on a unified state tuple $s_t$ containing end-effector pose, fingertip positions, calibrated forces ( $f_t$ ), spatial torque descriptors ( $\gamma_t$ ), and the shared pose latent ( $\psi_t$ ).
Structured Visual Masking: During training, visual tokens (patches or entire camera streams) are randomly replaced with learned mask embeddings. This forces the policy to rely on force and proprioception when visual evidence is missing or unreliable (e.g., during self-occlusion), rather than learning to ignore force data.
Training Objective: The policy predicts velocity vectors to transport a Gaussian noise source to a target action chunk, minimizing the flow-matching loss.

3. Hybrid Execution Pipeline

The system does not rely on a single end-to-end policy for long-horizon tasks. Instead, it composes learned primitives with model-based controllers:

Reach-and-Lift: Executed by MARC, handling the contact-sensitive phase of grasping and initial lifting.
Transport: Managed by a steady-grasp regulator that uses a hybrid force-position controller. It maintains a minimal stable grip and self-tightens based on calibrated torque feedback if transport loads the object.
Handover: Controlled by an end-effector admittance controller that yields the arm and opens the hand proportionally to external wrenches.
Controller Logic: The low-level controller tracks a commanded pose but applies a "one-sided retreat" correction based on calibrated torque. If contact torque exceeds a target, the controller opens the joint; otherwise, it follows the pose command. This ensures compliance without driving the hand deeper into the object.

Key Contributions

Transferable Contact Representation: A method to convert hand-specific effort signals into calibrated joint torques, fingertip forces, and compact load descriptors, making contact feedback physically comparable across different dexterous hands.
Unified Force-Position Interface: A framework combining shared hand-pose intent with calibrated contact feedback, enabling compliant control transfer across heterogeneous embodiments.
MARC Policy and Pipeline: A mask-aware flow-matching policy trained on vision, proprioception, and calibrated contact, integrated with model-based primitives for long-horizon manipulation (reach, transport, handover) without task-specific retraining.

Experimental Results

The system was evaluated on a Unitree G1 humanoid equipped with three structurally different hands (Wuji, Wuji12, and Inspire) across ten tasks involving rigid and compliant/deformable objects (e.g., eggs, brioche buns, stacked cups).

Performance Gain: MARC improved the mean reach-and-lift success rate from 0.44 (baseline sDiT) to 0.71 across all hands and tasks.
Specific Improvements:
- Inspire: 0.54 $\to$ 0.84
- Wuji: 0.48 $\to$ 0.79
- Wuji12: 0.29 $\to$ 0.51
- Critical Cases: The most significant gains were observed on the "Marker" task (slippery, occluded), where success rates jumped from 0.2/0.1 to 1.0/0.9 for Inspire and Wuji respectively.
Ablation: Removing the calibrated contact channel collapsed performance on rigid objects back to the baseline (0.48), confirming that the gain is driven by the force channel, not just architecture or masking.
Zero-Shot Generalization: The policy successfully transferred to an unseen Wuji15 configuration (15-DoF) without retraining, achieving success rates within 0.1 of the in-distribution Wuji results.

Significance and Claims

The paper claims that contact itself is transferable when expressed in physical units (torque and force) rather than hand-specific proxies. By unifying motion intent and contact loading into a shared interface, the system enables dexterous manipulation skills to transfer across structurally different hands. The authors emphasize that this approach allows for stable, compliant interaction over long-horizon tasks, particularly in scenarios where visual feedback is insufficient due to occlusion or object deformation. The work demonstrates that a single policy, trained on calibrated contact, can drive heterogeneous hardware without requiring per-hand model training or task-specific fine-tuning.