TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

The Torque Adaptation Module (TAM) is a novel learned module designed to achieve robust motion transfer in robot manipulation, addressing discrepancies arising from the sim-to-real gap, unknown payloads, or variations between robot instances. Operating between the low-level controller and the robot's torque interface, TAM comprises a history encoder that embeds proprioceptive history into a latent state and a torque adaptor that computes residual torque corrections. A key advantage is its ability to reuse the same TAM weights across policies with different action spaces (joint targets, end-effector targets, direct torques) without requiring policy retraining or domain randomization. TAM is trained entirely in randomized simulation, using multi-robot pretraining followed by robot-specific fine-tuning, requiring no real-robot data. Evaluated zero-shot on a real Franka Panda robot, TAM significantly improved performance in dynamic manipulation tasks, including a vision-based box pushing policy (RL), a flip policy (BC), and an MPC ball-on-plate balancing controller, outperforming online system identification and RMA baselines.

Key takeaway

For Robotics Engineers and ML Engineers deploying robot policies, if you are struggling with sim-to-real gaps, unknown payloads, or robot instance variations, consider integrating TAM. This module allows you to achieve robust real-world performance with existing policies without costly retraining or extensive real-robot data collection. You can enhance dynamic, contact-rich manipulation tasks by adding TAM as a modular, torque-level adaptation layer.

Key insights

TAM adapts torque commands at the robot interface, enabling robust motion transfer and decoupling policy training from robot dynamics.

Principles

• Adaptation at the torque interface enhances policy reusability.
• Proprioceptive history can infer robot dynamics mismatch.
• Sim-only training can yield effective zero-shot real-world transfer.

Method

TAM employs a history encoder to summarize proprioceptive history into a latent state and a torque adaptor to predict residual torque corrections at 1 kHz, applied between the low-level controller and the robot plant.

In practice

• Integrate TAM with existing RL, BC, or MPC policies.
• Fine-tune TAM in simulation for specific robot models.
• Apply to contact-rich dynamic manipulation tasks.

Topics

• Sim-to-Real Transfer
• Robot Manipulation
• Torque Control
• Adaptive Control
• Reinforcement Learning
• Behavior Cloning
• Model Predictive Control

Code references

• google-deepmind/mujoco_menagerie

Best for: Research Scientist, AI Scientist, Robotics Engineer, Machine Learning Engineer

Related on AIssential

• TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation — TAM enables robust robot policy transfer by adapting torques using proprioceptive history, trained purely in simulation.
• Sample-efficient Low-level Motion Planning for Robotic Manipulation Tasks via Zero-shot Transfer Learning — Leveraging transfer learning and reward redesign significantly improves robotic motion planning efficiency.
• A neural blueprint for human-like intelligence in soft robots — A neural-inspired AI control system enables soft robots to learn once and adapt instantly to diverse tasks and disturbances.
• v205: Proceedings of CoRL 2022 — Volume 205 of the Conference on Robot Learning, held in December 2022, showcases significant advancements across various facets of robot learning.
• Beyond Binary: Sim-to-Real Dexterous Manipulation with Physics-Grounded Contact Representation — Center-of-Pressure (CoP) enables robust sim-to-real dexterous manipulation by preserving dense, physics-grounded tactile information.
• Robotic Policy Adaptation via Weight-Space Meta-Learning — WIZARD enables robotic VLA policy adaptation via weight-space meta-learning, generating LoRA parameters from video and language without fine-tuning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.