TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation

Abstract: A policy tuned for one robot often behaves differently on another, whether due to the sim-to-real gap, unknown payloads, or the differing dynamics of two instances of the same robot. In contact-rich, dynamic manipulation, even small motion discrepancies can result in failure to track reference motion, since they disrupt the timing and modes of contact. Common remedies, such as domain randomization or system identification, either produce overly conservative task policies or require data that must be recollected for each robot or payload. We introduce the Torque Adaptation Module (TAM), a learned module that adapts the torque commands sent to the robot to match the behavior of an ideal robot. TAM operates between the low-level controller that tracks the policy's actions and the robot's torque interface. It includes a history encoder that embeds proprioceptive history into a latent state and a torque adaptor that computes residual torque corrections. Because TAM depends only on proprioceptive history and not on policy observations, or the action space, the same TAM weights can be reused to adapt policies with different action spaces (joint targets, end-effector targets, or direct torques). The policies themselves do not need to be trained with domain randomization of robot parameters. Instead, we offload the need for domain randomization to TAM by training it entirely in randomized simulation, using multi-robot pretraining followed by a robot-specific fine-tuning step that still requires no real-robot data. We evaluate TAM zero-shot on a real Franka Panda robot across dynamic manipulation tasks that include a vision-based box pushing policy (from RL), a flip policy (from BC), and an MPC ball-on-plate balancing. Our experiments show that TAM improves zero-shot real-robot execution compared to online system identification and RMA baselines and enables robust dynamic manipulation performance.

• The paper introduces TAM, a torque adaptation module that decouples adaptation from policy, correcting dynamics mismatches for robust sim-to-real transfer.
• It employs a long-horizon history encoder and a fast real-time torque adaptor to compute residual corrections below various control policies.
• Empirical results demonstrate significant improvements in success rates across RL, BC, and MPC tasks, validated on a Franka Panda and multiple robot models.

Torque Adaptation Module: Robust Sim-to-Real Transfer at the Torque Interface

Motivation and Problem Setting

Motion transfer for manipulation policies across simulation-to-real and robot-to-robot gaps remains fundamentally limited by drift in physical parameters, unknown exogenous effects (e.g., payload, wear), and low-level actuator/model mismatches. Conventional mitigation strategies such as domain randomization or model retuning introduce undesired conservatism and require either data collection for each robot or broad retraining. Policy-side adaptation frameworks, including RMA and UPOSI, condition policies on inferred robot-parameter embeddings, but cannot be decoupled from policy implementation—they require retraining or distillation for every new policy or change in action space.

The "TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation" (2606.06218) addresses this problem by introducing a learned dynamics adaptation module deployed at the torque interface, mediating between the low-level controller and the robot plant. TAM operates exclusively from proprioceptive history and applied torques, inferring unobserved dynamics mismatch and generating residual corrections to nominal torques at kHz rates, without access to policy observations or task reward.

TAM Approach and System Architecture

TAM comprises two core network modules: a long-horizon history encoder and a real-time torque adaptor. The first encodes a window of robot-side proprioceptive and torque history into a compressed latent state reflecting the current regime of physical mismatch; the second uses this latent and a short state window to compute the additive torque correction at each time step.

This design allows seamless attachment below various independently authored policies (RL, BC, MPC) with action spaces of joint targets, end-effector targets, or direct torque, assuming only that the low-level controller outputs a nominal torque.

Figure 1: TAM deployment interface supporting combinations of policies and low-level controllers, with TAM inserted at the torque interface to correct for dynamics mismatch in real time.

Crucially, TAM is trained entirely in simulation using randomized physics parameterizations and synthetic joint-space waypoint tracking. Supervisory signal is generated via analytically invertible simulated models, matching the local acceleration response of the ideal robot.

Figure 2: TAM data generation: randomized free-space joint trajectories with synthetically injected perturbations allow coverage of dynamics mismatches for effective local-supervision of the torque correction.

Training Objective and Features

During training, for any state and nominal torque, TAM’s supervision target is the torque needed by the perturbed robot to elicit the ideal-model acceleration response given identical external loading. This fully decouples the adaptation module from the downstream task or policy, concentrating all adaptation learning at the torque control layer.

A key architectural element is the inclusion of a "physics residual" feature in the encoder input: for each history step, the discrepancy between the applied torque and the torque required by the ideal robot (estimated via inverse dynamics) is provided. This gives the encoder a powerful direct signal for extracting robot-specific mismatch.

TAM splits deployment into two asynchronous rates: the history encoder runs at ~5Hz, updating the latent, while the torque adaptor runs at the 1kHz control rate, ensuring real-time deployment without GPU reliance in the control loop.

Empirical Evaluation

The evaluation covers dynamic real-world manipulation and tracking tasks on a Franka Panda manipulator, including:

• Vision-based RL policy for non-prehensile box pushing.
• Behavior-cloned visual cube-flipping (contact-rich).
• Ball-on-plate tracking with an MPC policy and unknown payloads.

In all cases, policies are obtained relative to the nominal ideal robot. TAM is deployed at the torque interface on the real robot, which is subject to unknown payloads or physical parameter shifts.

Figure 3: Real-world execution examples: comparison between direct policy execution (no adaptation) and TAM-enabled execution for RL box pushing, BC cube flipping, MPC ball-balancing, and payload tracking.

Notable numerical results include:

• On real-world RL box pushing, TAM attains a 76.2% success rate, outperforming direct transfer (47.6%), online system identification (57.1%), and RMA adaptation in the policy (52.4%).
• On BC flip, TAM achieves 72% success versus 50% (direct) and 34% (online SysID), closely matching the ideal-simulation reference (73.5%).
• For MPC ball-on-plate, TAM matches online SysID performance and approaches the ideal policy upper bound.

TAM’s adaptation remains robust under both low-level controller switching (joint impedance to operational-space control) and significant actuation changes (1kg end-effector payload), yielding the lowest tracking RMSE in every configuration.

Figure 4: End-effector tracking of carried-state TAM versus direct and SysID-based adaptation after controller switch and payload addition.

Furthermore, a cross-robot generalization study using seven robot models in simulation shows TAM’s multi-robot pretraining yields consistent joint-tracking RMSE reduction, even on unseen robots. Fine-tuning with only simulated trajectories on the target yields further performance gains.

Analysis and Ablations

The paper reports extensive ablation studies: removing the history encoder or the physics residual feature consistently degrades transfer, emphasizing their impact. Directly predicting explicit physical parameters from history and applying analytic torque adaptation underperforms TAM’s latent-to-torque approach, validating the choice of amortizing all mismatch into the network’s latent bottleneck. Latent-state analysis (t-SNE) confirms that the learned embedding organizes primarily by physical condition rather than trajectory identity, indicating effective invariance and factorization.

Figure 5: t-SNE embedding of history token features—coloring by trajectory ID (left) and physical condition (right) highlights that physical regime is the dominant organizing factor.

Limitations

TAM assumes access to a torque interface and structural consistency in robot kinematics and actuator ordering; it cannot compensate mismatch outside its hardware torque limits and does not distinguish robot-plant mismatch from unexpected large external loads in deployment. The adaptation is fundamentally local: long-term stability and safety rely on sufficient coverage of the operation space during synthetic training and appropriate policy-level constraints.

Implications and Future Directions

TAM delivers a torque-residual adaptation abstraction at a physically universal interface, decoupling the adaptation module from the specifics of policy implementation or supervision. This allows the deployment of learned, behavior-cloned, or model-based policies optimized solely for the ideal robot, bypassing the necessity of task-domain randomization or repeated explicit identification.

Practically, widespread adoption of plug-and-play torque adaptation modules could facilitate scalable robot fleet deployment, reduce tuning effort, and enable robust cross-vendor or cross-instance transfer of manipulation skills.

Theoretically, this approach highlights the potential of history-encoded, bottlenecked latent representations at intermediate control layers, distinct from explicit symbolic parameter identification. Future extensions could incorporate force/torque sensor streams, address richer contact-manipulation phenomena, or generalize the architecture for non-kinematically matched platforms.

Conclusion

The paper introduces and validates the Torque Adaptation Module (TAM) as an effective, modular, and policy-agnostic mechanism for robust sim-to-real and cross-robot manipulation transfer. Systematic empirical validation demonstrates superior zero-shot real-world performance compared to policy-side adaptation and explicit system identification methods. This decoupling of adaptation from policy will likely inform future robust robot learning workflows and transfer frameworks for high-performance control (2606.06218).