ECHO

ECHO: In-Context Behavior Foundation Model
for Humanoid Whole-Body Control

Xingyi Wang1,2*  ·  Bin Li3*  ·  Jiyuan Shi2  ·  Chenyun Zhang2  ·  Weiji Xie2,4  ·  Xiangyu Miao2,4  ·  Junsong Wu2,4  ·  Wei Song1  ·  Shiqiang Zhu1†  ·  Chenjia Bai2†

1 Zhejiang University
2 Institute of Artificial Intelligence (TeleAI), China Telecom
3 ShanghaiTech University
4 Shanghai Jiao Tong University

* Equal contribution (co-first authors). † Corresponding authors.

ECHO Overview

ECHO pairs a broadly trained whole-body behavior prior with a context-conditioned residual controller.Recent proprioceptive and action history informs adaptation to changing physical dynamics, without test-time parameter updates.

Method Pipeline

A broad motion-tracking prior supplies the skill; interaction context adjusts how that skill is executed.

Figure 2 method pipeline: MIAS behavior-prior pretraining, context representation learning, and residual adaptation.
Figure 2 · Method overview. The diagram is cropped from the manuscript. Snowflakes mark frozen modules during the relevant stage; flames mark trainable modules.
01 · MIAS

Build the behavior prior

Mixed and Annealed Sampling balances targeted retries of difficult motions with broad coverage of the motion corpus while pretraining ECHO-PRE.

02 · CONTEXT LEARNING

Represent physical context

A context encoder learns from state-action histories using dynamics prediction and physical-neighborhood soft alignment.

03 · RESIDUAL ADAPTATION

Correct the prior's action

A residual actor uses the learned context to adjust the frozen prior. The prior and encoder remain frozen during residual-policy training.

Experimental Results

Motion-tracking results from the manuscript. SR is success rate; arrows show the preferred direction.

Motion Tracking Fidelity

MIAS improves the pretrained policy on both everyday and high-dynamic motions; CRA further improves tracking precision under nominal dynamics.

MethodSR ↑Eg-MPJPE ↓EMPJPE ↓Eacc ↓Evel ↓
AMASS-2000-Random
SONIC v1.10.9633 ± 0.0003316.651 ± 1.64930.411 ± 0.0151.135 ± 0.0023.770 ± 0.006
HEFT0.9785 ± 0.0009261.573 ± 2.35228.524 ± 0.0091.062 ± 0.0013.103 ± 0.004
ECHO-PRE w/o MIAS0.9720 ± 0.0005331.235 ± 2.13232.580 ± 0.0181.243 ± 0.0034.170 ± 0.008
ECHO-PRE0.9832 ± 0.0012207.995 ± 0.97026.595 ± 0.0070.932 ± 0.0033.036 ± 0.006
ECHO-N0.9827 ± 0.0003203.149 ± 0.28925.577 ± 0.0240.917 ± 0.0032.956 ± 0.007
BONES-SEED-100-HighDynamic
SONIC v1.10.5800 ± 0.0100343.124 ± 4.42161.791 ± 2.6154.086 ± 0.03110.406 ± 0.102
HEFT0.6200 ± 0.0173391.094 ± 30.75859.562 ± 0.1353.232 ± 0.0178.229 ± 0.138
ECHO-PRE w/o MIAS0.4500 ± 0.0100365.075 ± 5.27064.580 ± 0.0254.297 ± 0.04211.712 ± 0.145
ECHO-PRE0.7200 ± 0.0000244.772 ± 3.14356.521 ± 0.3352.801 ± 0.0127.507 ± 0.034
ECHO-N0.7633 ± 0.0115235.410 ± 4.14055.392 ± 0.4152.766 ± 0.0057.360 ± 0.017

Table 2 · Mean ± sample standard deviation over three runs with independent action-noise seeds (σ = 0.01). AMASS-2000-Random tests everyday motions excluded from training; BONES-SEED-100-HighDynamic tests challenging motions within the training distribution.

Takeaway. On the high-dynamic benchmark, ECHO-N reaches 0.7633 SR versus 0.4500 for the same prior trained without MIAS.

Context Representation

Interaction history reveals the physical environment and informs adaptation.

Figure 3 t-SNE plots: ECHO forms clearer environment-wise clusters than Any2Track.
Figure 3 · Context representation comparison. t-SNE projections colored by environment. ECHO shows clearer environment-wise grouping across motions than Any2Track.
Figure 4 plots: environment accuracy, cross-motion retrieval mAP, and payload mass R squared rise with context length.
Figure 4 · Context-length scaling. Longer histories improve environment identification, cross-motion retrieval, and payload-mass readout on LAFAN-40. Lines show means; shading shows sample standard deviation over three evaluations.

Demonstration Gallery

Explore acrobatics, ground and terrain interaction, and in-context adaptation.