ZEST: Zero-shot Embodied Skill Transfer for Athletic Robot Control

TL;DR

ZEST tackles robust, human-like whole-body control for humanoid robots by learning from diverse motion data and deploying zero-shot to hardware. It uses a single-stage reinforcement learning pipeline trained from MoCap, ViCap, and keyframe animation without contact labels or state estimators, and deploys directly to Atlas, G1, and Spot. The framework introduces a model-based virtual assistive wrench to stabilize long-horizon motions and an adaptive RSI sampling strategy to focus data collection on difficult segments, along with progressive PLA actuator modeling to improve sim-to-real transfer. Together, these contributions enable dynamic, multi-contact behaviors on a full-size humanoid and cross-morphology transfers, outperforming a whole-body MPC baseline in robustness and flexibility, while highlighting areas for generalization and perception-enabled extensions.

Abstract

Achieving robust, human-like whole-body control on humanoid robots for agile, contact-rich behaviors remains a central challenge, demanding heavy per-skill engineering and a brittle process of tuning controllers. We introduce ZEST (Zero-shot Embodied Skill Transfer), a streamlined motion-imitation framework that trains policies via reinforcement learning from diverse sources -- high-fidelity motion capture, noisy monocular video, and non-physics-constrained animation -- and deploys them to hardware zero-shot. ZEST generalizes across behaviors and platforms while avoiding contact labels, reference or observation windows, state estimators, and extensive reward shaping. Its training pipeline combines adaptive sampling, which focuses training on difficult motion segments, and an automatic curriculum using a model-based assistive wrench, together enabling dynamic, long-horizon maneuvers. We further provide a procedure for selecting joint-level gains from approximate analytical armature values for closed-chain actuators, along with a refined model of actuators. Trained entirely in simulation with moderate domain randomization, ZEST demonstrates remarkable generality. On Boston Dynamics' Atlas humanoid, ZEST learns dynamic, multi-contact skills (e.g., army crawl, breakdancing) from motion capture. It transfers expressive dance and scene-interaction skills, such as box-climbing, directly from videos to Atlas and the Unitree G1. Furthermore, it extends across morphologies to the Spot quadruped, enabling acrobatics, such as a continuous backflip, through animation. Together, these results demonstrate robust zero-shot deployment across heterogeneous data sources and embodiments, establishing ZEST as a scalable interface between biological movements and their robotic counterparts.

Figures (10)

• Figure 1: Hardware deployment of ZEST across diverse data sources and robot morphologies. In order of appearance from top left to bottom right, the figure illustrates the following behaviors. From MoCap: Crawl on all fours (Atlas), roll on all fours (Atlas), jog (Atlas), breakdance (Atlas), forward roll (Atlas), cartwheel (G1), table-tennis (G1), cartwheel (atlas), army crawl (Atlas). From ViCap: Dance snippet A (Atlas), jump onto box (G1), climb up/down box (G1), ballet (G1), dance snippet C, soccer kick (Atlas). From Animation: handstand invert (Atlas), handstand balance (Spot), continuous backflip (Spot), barrel roll (Spot).
• Figure 2: Summary of the ZEST framework and its hardware results.
• Figure 3: Simulation-based evaluation and ablation studies. (a–d) A mid-training snapshot of the signals driving our adaptive curriculum: (a) assistive wrench scaling, which is modulated by failure, (b) per-bin failure levels, (c) bin visitation counts, and (d) sampling probabilities. Higher failure rates increase the sampling probability, and over time, visitation counts rise where failures are frequent. (e–f) Success rates for the baseline and ablations after 10 h and 20 h of training under complete domain randomization. Bars show mean success; whiskers show min–max; internal lines mark median; dots denote p10; numbers under lower whiskers report the minimum. The plots show that the assistive curriculum and adaptive sampling are critical for performance and sample efficiency. Removing privileged information from the critic or using absolute actions significantly degrades robustness, while adding observation/reference windows hinders learning.
• Figure 4: Overview of ZEST, which consists of three main stages.(1) Reference data: A diverse set of motions from MoCap, ViCap, and keyframe animation is processed; MoCap/ViCap references are kinematically retargeted to the target robot. (2) Training setup and MDP formulation: In simulation, the policy is trained using only on-robot signals and the next target state from the reference, while a separate critic receives privileged information (e.g., true base velocity, contact forces) to accelerate learning. To handle long-horizon clips and scale beyond single skills, an adaptive sampling scheme is proposed: trajectories are segmented into fixed-duration bins; a per-bin difficulty level metric is updated via an EMA of failure scores; and a categorical sampler biases reset-state selection toward harder bins while avoiding catastrophic forgetting of easier behaviors. A model-based virtual assistive wrench is applied at the base to stabilize training for highly dynamic behaviors; the current bin's difficulty level modulates its magnitude and is automatically annealed to zero as tracking improves. (3) Zero-shot deployment: The trained policy is deployed directly to the physical robot without any fine-tuning.
• Figure 5: Progressive simplification of proposed Parallel-Linkage Actuator models. It illustrates our modeling approach using a representative humanoid leg, where a motor on the thigh actuates the knee via a four-bar linkage. We begin with (1) the Exact Model, which fully resolves the closed-loop dynamics. We then introduce a series of progressively simplified models. (2) The Locally Projected Model assumes massless support links. (3) The Dynamic/Nominal Armature Model uses a Jacobi approximation for coupled joints, and finally, the most efficient model calculates armature values at a single fixed configuration.