ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting

ReActor's key novelty is the bilevel formulation itself: rather than retarget motion first then train a policy, it jointly optimizes morphology adaptation and policy learning. This avoids the compounding error problem where naive retargeting produces physically invalid trajectories that poison downstream imitation learning.

This work sits in the same reinforcement learning + robotics lineage as SAVGO (the cosine similarity method from early May), which also tackled sample efficiency in continuous control by rethinking how value information shapes action selection. Where SAVGO focused on representation geometry, ReActor focuses on the upstream problem: ensuring the demonstration data itself is physically coherent before policy learning begins. Both papers address the gap between what gradient-based RL can theoretically do and what it can actually do with real robot constraints. The Meta acquisition of Assured Robot Intelligence signals that embodied AI infrastructure is becoming a platform play, and motion retargeting is exactly the kind of foundational capability that platforms need to commoditize.

If ReActor's approach generalizes to morphologies significantly different from the training distribution (e.g., quadrupeds trained on bipedal mocap), that validates the claim about reducing manual engineering. If the method requires per-morphology hyperparameter tuning despite the automatic tuning claim, the deployment friction advantage evaporates.