ECHO: In-Context Behavior foundation model for Humanoid Whole-Body Control
TL;DR ECHO pairs one broad motion-tracking policy for the Unitree G1 with a residual policy that reads the robot's recent states and actions to adapt to payloads and other dynamics changes, with no test-time weight updates. It raises simulated success on high-dynamic motions to 0.76 (HEFT 0.62) and has the lowest real-robot tracking error on all four test motions, with or without 2 kg on each lower leg.
What they do
Humanoid behavior foundation models can track a huge range of reference motions. But a single policy trained across randomized dynamics tends to settle on conservative compromises when payload, friction or actuator properties change. ECHO splits the problem in two: a broad motion-tracking prior, and a residual policy that infers the current dynamics from the robot’s own recent states and actions and corrects the prior’s actions. Nothing is retrained at test time. All experiments use the Unitree G1, and on hardware it is driven by real-time teleoperation.
How it works
- Behavior prior (ECHO-PRE). PPO motion tracking in Mjlab on about 1,236 hours of filtered motion data, with AMASS held out for testing. Mixed and Annealed Sampling restarts training at a failed segment after a failure, and otherwise samples segments uniformly or by failure rate. The focus on hard segments is relaxed over training.
- Context encoder. Reads a 1 s short-term history plus a memory of earlier chunks from the same physical session. It is trained on the frozen prior’s rollouts to predict future states and contacts over several steps. A soft-alignment loss keeps environments with similar physical parameters close in latent space across different motions.
- Residual adaptation. With the prior and encoder frozen, an RL residual actor adds context-conditioned corrections. ECHO-H is trained under heavy randomization with up to 2.5 kg per hand and 4 kg per shin.
Results
- Simulated tracking of 100 high-dynamic motions: success goes from 0.45 with uniform sampling to 0.72 with MIAS and 0.76 with the residual, against 0.62 for HEFT and 0.58 for SONIC v1.1.
- Under heavy randomization ECHO-H reaches 0.892 success, against 0.883 for an RMA teacher that sees the true dynamics parameters and 0.870 for the RMA student.
- The context gives 97.5% environment-identification accuracy and linear-readout R² of 0.84 for payload mass and 0.92 for friction. Without the prediction loss, friction R² falls to -0.01.
- Real G1, 10 trials per motion: without payload, HEFT fails sit-down/stand-up and prone-crawl/stand-up, and SONIC fails sit-down, while ECHO completes all four. With 2 kg on each lower leg, ECHO-H has the lowest error on every motion (Charleston 52 mm vs 64 mm for the no-context residual). With 10 kg spread over its hands and legs, it still balances on one leg and squats.
Limitations
The context is learned only from simulated rollouts with privileged supervision, so effects outside the training randomization are untested. Memory assumes fixed physics within a session, which leaves adaptation to unannounced mid-session changes largely unexplored. Only the G1 and reference-conditioned tracking are evaluated.
Videos
Figures




Abstract
Humanoid behavior foundation models (BFMs) acquire diverse skills through large-scale motion imitation and massively parallel reinforcement learning, yet broad behavioral capability does not ensure precise execution across physical conditions. Without distinguishing the underlying dynamics, policies may favor conservative compromises over environment-specific precision. We introduce ECHO, an in-context BFM that turns interaction experience into environment-specific control, separating general behavioral capability from dynamics-specific execution. Mixed and Annealed Sampling first establishes a behavior prior in a single pretraining stage by balancing targeted practice with broad motion coverage. Building on this prior, Context Representation and Residual Adaptation learns structured dynamics representations from interaction histories to guide residual corrections to the frozen prior, enabling adaptation without test-time parameter updates. Extensive experiments show that ECHO outperforms all baseline methods across diverse motions and physical conditions. Moreover, deployment on the Unitree G1 demonstrates whole-body teleoperation spanning extreme acrobatics and terrain interaction, alongside payload adaptation in manipulation and whole-body control. Page is available at https://mwondering.github.io/echo-bfm/
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
