TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation

2026-06-04 · Source: Artificial Intelligence · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The Torque Adaptation Module (TAM) is a learned system designed to improve robust motion transfer in dynamic, contact-rich robot manipulation. It addresses challenges like the sim-to-real gap, unknown payloads, and differing robot dynamics that cause policies to fail when transferred between robots. TAM functions by adapting torque commands between the low-level controller and the robot's interface, utilizing a history encoder for proprioceptive data and a torque adaptor for residual corrections. A key advantage is its independence from policy observations or action spaces, allowing the same TAM weights to be reused across policies with joint, end-effector, or direct torque targets. Policies themselves do not need domain randomization; TAM is trained entirely in randomized simulation through multi-robot pretraining and robot-specific fine-tuning, requiring no real-robot data. Experiments on a real Franka Panda robot showed TAM improved zero-shot execution for tasks including vision-based box pushing, a flip policy, and MPC ball-on-plate balancing, surpassing online system identification and RMA baselines.

Key takeaway

For Robotics Engineers developing manipulation policies, if you face challenges with sim-to-real gaps or transferring policies across different robot instances or payloads, you should consider integrating a Torque Adaptation Module (TAM). This approach allows your policies to achieve robust zero-shot real-robot execution in dynamic, contact-rich tasks. By training TAM entirely in randomized simulation, you can significantly reduce the need for real-robot data collection and avoid policy-specific domain randomization, streamlining your development workflow.

Key insights

TAM enables robust robot policy transfer by adapting torques using proprioceptive history, trained purely in simulation.

Principles

• Proprioceptive history enables policy-agnostic adaptation.
• Simulated domain randomization can replace real-world data collection.
• Residual torque corrections enhance motion tracking robustness.

Method

TAM operates between the low-level controller and robot torque interface. It uses a history encoder for proprioceptive data and a torque adaptor for residual torque corrections. Training involves multi-robot pretraining and robot-specific fine-tuning in randomized simulation.

In practice

• Reuse TAM weights across diverse robot action spaces.
• Train adaptation modules entirely in randomized simulation.
• Apply TAM for dynamic, contact-rich manipulation tasks.

Topics

• Robotics
• Manipulation
• Torque Control
• Sim-to-Real Transfer
• Reinforcement Learning
• Franka Panda

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

Related on AIssential

• TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation — TAM adapts torque commands at the robot interface, enabling robust motion transfer and decoupling policy training from robot dynamics.
• FACTR 2: Learning External Force Sensing for Commodity Robot Arms Improves Policy Learning — FACTR 2 enables force-aware robotics on commodity arms by estimating torques and optimizing policy learning without extra sensors.
• Blind Dexterous Grasping via Real2Sim2Real Tactile Policy Learning — Calibrated simulation and geometry-aware tactile representations enable blind dexterous grasping on physical robots.
• Japan Science and Technology Agency Develops NVIDIA-Powered Moonshot Robot for Elderly Care — Robots can predict 3D interaction forces from RGB images using simulation-trained AI.
• 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.
• Mana: Dexterous Manipulation of Articulated Tools — Dexterous manipulation of articulated tools can be effectively modeled and solved as an animation problem using a sim-to-real framework.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.