Stubborn: A Streamlined and Unified Reinforcement Learning Framework for Robust Motion Tracking and Fall Recovery for Humanoids

A research team has proposed Stubborn, a reinforcement learning framework designed to unify robust motion tracking and fall recovery for humanoid robots, addressing a key limitation in current methods that treat the two capabilities as separate tasks [1]. The framework, detailed in a paper submitted to arXiv in 2026, moves away from conventional approaches that require multi-stage training with specialized recovery rewards or separate policies [1][2]. Existing reinforcement learning methods often terminate training episodes immediately after severe tracking failures, which the authors argue restricts recovery-oriented exploration when a robot is unstable or has fallen [1][2]. Stubborn employs an asymmetric Actor-Critic architecture and introduces three core components to overcome these constraints [1][2]. A yaw-aligned tracking representation is used to reduce sensitivity to global drift and heading disturbances while preserving gravity-related balance information [1][2]. The framework also incorporates a Bernoulli-based probabilistic termination mechanism, which encourages the policy to explore fall-recovery behaviors across varying failure modes instead of halting the episode [1][2]. A third element, a probabilistic termination and tracking-error-driven strategy, dynamically reshapes the sampling distribution based on tracking performance to increase training efficiency for difficult motion segments and unstable states [1][2]. The authors report that extensive comparisons with state-of-the-art methods and ablation studies showed Stubborn achieved competitive performance, with the probabilistic termination mechanism and adaptive sampling strategy contributing to performance and robustness gains [1][2]. Real-world demonstrations of the system are available on the project's website [1][2]. The work was posted on the arXiv preprint server, a common venue for early-stage computer science and robotics research, and associated tooling links reference platforms such as Hugging Face for potential model and data sharing [1].