TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation

Researchers introduce Torque Adaptation Module (TAM), a learned module that adapts robot torque commands to compensate for dynamics differences across robot instances, payload variations, and sim-to-real gaps. TAM enables reusable policy adaptation without requiring robot-specific retraining or real-world data collection, demonstrating robust performance on dynamic manipulation tasks with a real Franka Panda robot.

TAM addresses a fundamental challenge in robotics: policies trained for one robot configuration often fail when deployed on different hardware or with different payloads. Traditional solutions like domain randomization require overly conservative policies, while system identification demands expensive data collection for each new robot instance. This research offloads adaptation responsibilities to a separate learned module operating at the torque level, enabling policy reuse across different action spaces and robot configurations.

The technical innovation centers on TAM's architecture: a history encoder that embeds proprioceptive data into latent representations, coupled with a torque adaptor computing residual corrections. Critically, TAM depends only on proprioceptive history—joint positions, velocities, and torques—rather than policy observations or specific action spaces. This design choice enables remarkable flexibility: the same TAM weights generalize across policies using joint targets, end-effector targets, or direct torque commands. Training occurs entirely in simulation with multi-robot pretraining, followed by minimal robot-specific fine-tuning without requiring real-world data.

For the robotics and automation industry, TAM represents significant practical value. Manufacturers deploying manipulation systems across multiple robot instances can now achieve consistent performance without extensive retraining pipelines. The approach reduces operational costs by eliminating per-robot data collection requirements while maintaining task robustness in contact-rich dynamics where timing and contact modes are critical. Successful zero-shot evaluation across diverse tasks—vision-based pushing, flipping, and balancing—demonstrates TAM's generalization capability. This work suggests a broader trend toward decoupling policy learning from system-specific adaptation, potentially accelerating robotic system deployment in industrial and research settings.

• →TAM enables single-policy deployment across multiple robot instances and payload configurations without real-world retraining data.
• →The module operates at the torque interface level, depending only on proprioceptive history rather than policy observations or action spaces.
• →Multi-robot pretraining followed by minimal fine-tuning eliminates the need for expensive per-robot data collection.
• →Zero-shot evaluation demonstrates robust performance on diverse manipulation tasks including RL policies, behavioral cloning, and MPC approaches.
• →TAM's design enables weight reuse across policies with different action space representations, maximizing deployment flexibility.