Toronto Researchers' TAM Fixes Robot Sim-to-Real Failures Without Retraining Policies
A new learned module corrects torque commands at runtime, letting any manipulation policy transfer to a real robot without retraining or real-world data collection.
Takeaways
- What it is: TAM is a learned module that corrects torque commands at runtime to compensate for sim-to-real gaps, unknown payloads, and robot-to-robot variation.
- Policy-agnostic: TAM wraps any policy (RL, BC, MPC) regardless of action space -- joint targets, end-effector targets, or direct torques -- without modifying the policy.
- No real-robot data needed: TAM is trained entirely in simulation using multi-robot pretraining followed by robot-specific fine-tuning, both simulation-only.
- Outperforms baselines: Zero-shot on a real Franka Panda, TAM beats online system identification and RMA across dynamic tasks including box pushing, cube flipping, and ball-on-plate balancing.
- Key design: TAM sits between the low-level controller and the robot's torque interface, using only proprioceptive history -- never policy observations or action spaces.
- No code release: The paper is on arXiv; no weights or code have been released alongside the publication.
Every roboticist knows the pain: you train a policy in simulation, deploy it on the real robot, and it falls apart. The motors behave slightly differently, the payload shifts the dynamics, or a second unit of the "same" robot just feels different. A new paper from researchers at the University of Toronto, McGill University, and the University of Washington introduces TAM (Torque Adaptation Module), a lightweight learned module that sits between your policy and the robot's motors and quietly corrects for all of that -- without touching the policy itself.
The gap that breaks dynamic manipulation
The sim-to-real problem is well-known, but it hits especially hard in contact-rich, dynamic manipulation. Reinforcement learning where a robot is trained in a simulated environment has made great progress, but suffers from the sim-to-real problem -- real motors and hardware do not exactly behave like the simulated ones. For slow, quasi-static tasks like pick-and-place, small discrepancies are forgivable. But for dynamic tasks -- flipping a cube, balancing a ball on a plate, or pushing a box at speed -- even a tiny timing error disrupts the contact sequence and causes failure.
The standard fixes each come with a cost. Domain randomization (training the policy across a wide range of simulated physics parameters to make it robust) trains the policy across a diverse range of simulated environments by varying parameters that influence the environment dynamics, such as physical properties like mass and friction coefficients, ensuring the policy is exposed to a broad spectrum of variations. But this tends to produce overly conservative policies that sacrifice peak performance for robustness. System identification -- measuring the real robot's parameters and plugging them into the simulator -- requires data collection every time you change robots or add a payload. Neither approach scales gracefully.