- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
ECHO combines a pretrained humanoid motion tracker with context-conditioned residual adaptation. It introduces MIAS for balancing broad motion coverage and difficult-segment practice, and CRA for learning dynamics representations from interaction histories. The framework is evaluated through simulation benchmarks, representation analyses, and real-world Unitree G1 demonstrations.
Critical strengths
The real-world demonstrations are impressive, covering highly dynamic motions, ground interactions, terrain traversal, and payload-conditioned behaviors with a shared control framework.
Separating broad behavioral capability from dynamics-specific execution is well motivated. The frozen-prior and residual-policy design provides a practical mechanism for adaptation without test-time parameter updates.
The evaluation includes tracking performance, adaptation baselines, representation analyses, and hardware experiments. Sharing the frozen backbone across adaptation baselines is a useful experimental control.
Critical weaknesses
-
The connection between the strongest demonstrations and the central in-context learning contribution remains unclear. The broad terrain and acrobatic demonstrations in Figure 6 are produced by ECHO-PRE, before introducing CRA. These results establish the strength of the behavior prior, but provide limited evidence for the necessity of the proposed adaptation mechanism. Although CRA is also evaluated in simulation, its direct real-world validation is concentrated on payload changes.
-
The contribution relative to existing BFMs is insufficiently established. The paper does not quantitatively compare with BFM, ScaleBFM, BFM-Zero, or Perceptive BFM. While these methods address different aspects of general-purpose control, a clearer comparison of capabilities, interfaces, and adaptation mechanisms is needed. The RMA and Any2Track comparisons are relevant to dynamics adaptation, but do not resolve the broader positioning of ECHO as a behavior foundation model.
-
The evidence connecting structured context representations to improved control is incomplete. The representation ablations and context-length experiments primarily measure environment identification, retrieval, and physical-parameter readout. They do not directly establish how the proposed representation objectives or additional context improve downstream tracking. Adaptation speed, cold-start behavior, and sensitivity to outdated context are also insufficiently quantified. Consequently, the necessity and practical benefit of the specific in-context learning design remain difficult to assess.
Questions for authors
-
Which demonstrated capabilities specifically depend on CRA? Could ECHO-PRE and the full ECHO model be evaluated on identical real-world motions and physical conditions to separate the contributions of the behavior prior and contextual adaptation?
-
Could the authors report downstream tracking success and errors after removing dynamics prediction or physical-neighborhood alignment? How does the full method compare with direct history-conditioned control under matched history length, model capacity, and training budget?
-
How does control performance improve as interaction context accumulates? What is the adaptation time from an empty history, and how is accumulated memory handled when a payload is attached or removed within a session?
-
How does ECHO compare with BFM, ScaleBFM, BFM-Zero, and Perceptive BFM in terms of control interfaces and demonstrated capabilities? Could matched quantitative comparisons be provided where applicable, with explicit discussion of differences in scope and assumptions?
10 comments
Table list (Table1-3)
We have added 300 highly dynamics motions on the original BONES-SEED-100-HighDynamic to compose BONES-SEED-400-HighDynamic, maintaining the same ratio: 1/3 terrain interaction, 1/3 extreme movements, and 1/3 dance behaviors. What's more, we added ScaleBFM, BFM-Zero and P-BFM as baselines. As illustrated in Table 1, ECHO-N achieved the best overall performance.
Table 1
Motion-tracking evaluation on BONES-SEED-400-HighDynamic. Results are reported as mean sample standard deviation over three independent action-noise seeds with . Bold and underlined values indicate the best and second-best means within each environment, respectively.
Nominal Environment
| Method | SR ↑ | ↓ | ↓ | ↓ | ↓ |
|---|---|---|---|---|---|
| SONIC v1.1 | 0.8750 ± 0.0043 | 383.362 ± 3.344 | 43.706 ± 0.519 | 3.316 ± 0.011 | 8.869 ± 0.027 |
| HEFT | 0.8833 ± 0.0029 | 270.710 ± 1.608 | 39.906 ± 0.051 | 2.929 ± 0.016 | 7.133 ± 0.027 |
| SCALEBFM-XL | 0.9275 ± 0.0050 | 249.122 ± 0.968 | ± 0.036 | 2.657 ± 0.002 | 6.329 ± 0.010 |
| BFM-ZERO | 0.2858 ± 0.0151 | 813.942 ± 15.446 | 152.798 ± 1.287 | 3.006 ± 0.053 | 13.555 ± 0.189 |
| P-BFM | 0.2475 ± 0.0090 | 653.343 ± 20.239 | 126.544 ± 4.389 | 4.576 ± 0.020 | 13.536 ± 0.229 |
| ECHO-PRE | 0.9200 ± 0.0000 | 218.547 ± 1.829 | 39.082 ± 0.109 | ± 0.010 | 6.827 ± 0.017 |
| ECHO-N | ± 0.0038 | ± 1.385 | 38.118 ± 0.068 | 2.594 ± 0.008 | ± 0.018 |
NDR Environment
| Method | SR ↑ | ↓ | ↓ | ↓ | ↓ |
|---|---|---|---|---|---|
| SONIC v1.1 | 0.8417 ± 0.0029 | 416.569 ± 4.220 | 45.345 ± 0.178 | 3.475 ± 0.003 | 9.341 ± 0.015 |
| HEFT | 0.8633 ± 0.0038 | 339.039 ± 3.282 | 42.549 ± 0.074 | 3.047 ± 0.011 | 7.580 ± 0.018 |
| SCALEBFM-XL | 0.8975 ± 0.0025 | 279.684 ± 4.908 | ± 0.127 | 2.781 ± 0.016 | 6.744 ± 0.046 |
| BFM-ZERO | 0.2492 ± 0.0063 | 870.781 ± 12.997 | 159.128 ± 3.725 | 2.903 ± 0.045 | 13.160 ± 0.247 |
| P-BFM | 0.1900 ± 0.0050 | 603.520 ± 13.907 | 127.483 ± 7.200 | 4.749 ± 0.058 | 13.851 ± 0.305 |
| ECHO-PRE | ± 0.0000 | ± 1.775 | 41.956 ± 0.039 | ± 0.001 | 7.188 ± 0.006 |
| ECHO-N | 0.9167 ± 0.0063 | 236.017 ± 1.410 | 40.342 ± 0.062 | 2.655 ± 0.004 | ± 0.020 |
Table 2
Comparison of and ECHO and ECHO-E2E. ECHO-E2E and ECHO share the same policy architecture, history capacity and training budget.
| Metric | ECHO-E2E | ECHO |
|---|---|---|
| Policy parameters | 19.150 M | 19.150 M |
| Training budget (environment steps) | 3,932,160,000 | 3,932,160,000 |
| History capacity (steps) | 350 (7 s at 50 Hz) | 350 (7 s at 50 Hz) |
Table 3
Motion-tracking evaluation on LAFAN40. Results are reported as mean sample standard deviation over three seeds, with action-noise standard deviation . Bold values indicate the best means within each environment.
Nominal Environment
| Method | SR ↑ | ↓ | ↓ | ↓ | ↓ |
|---|---|---|---|---|---|
| ECHO-E2E | 0.4083 ± 0.0144 | 5299.772 ± 181.732 | 168.213 ± 4.066 | 3.262 ± 0.006 | 19.831 ± 0.091 |
| ECHO-H | 0.5667 ± 0.0144 | 2907.805 ± 43.403 | 30.969 ± 0.034 | 2.801 ± 0.001 | 7.410 ± 0.008 |
HDR Environment
| Method | SR ↑ | ↓ | ↓ | ↓ | ↓ |
|---|---|---|---|---|---|
| ECHO-E2E | 0.4021 ± 0.0072 | 5586.404 ± 151.949 | 180.759 ± 3.080 | 3.322 ± 0.074 | 19.845 ± 1.108 |
| ECHO-H | 0.4646 ± 0.0095 | 3394.704 ± 160.671 | 39.016 ± 0.769 | 2.986 ± 0.033 | 8.761 ± 0.169 |
Table list (Table4-5)
Table 4
Simulation evaluation of BFMs' in-context ability. ECHO-H effectively exploits interaction context to improve tracking under heavily randomized dynamics (HDR). Bold and underlined values indicate the best and second-best results within each environment, respectively.
Nominal Environment
| Method | SR ↑ | ↓ | ↓ | ↓ | ↓ |
|---|---|---|---|---|---|
| Vanilla | 0.8892 ± 0.0003 | 194.541 ± 1.585 | 27.286 ± 0.016 | 0.881 ± 0.008 | 2.887 ± 0.006 |
| Any2Track | 0.8743 ± 0.0008 | 193.623 ± 0.738 | 27.846 ± 0.002 | 0.873 ± 0.008 | 2.883 ± 0.005 |
| RMA-Teacher | ± 0.0003 | 189.908 ± 0.220 | ± 0.010 | ± 0.007 | ± 0.004 |
| RMA-Student | 0.9763 ± 0.0008 | 190.329 ± 0.590 | 24.026 ± 0.007 | 0.851 ± 0.008 | 2.774 ± 0.005 |
| ECHO-PO | 0.9823 ± 0.0010 | 194.255 ± 0.344 | 25.125 ± 0.004 | 0.90597 ± 0.00008 | 2.898 ± 0.001 |
| ECHO-SO | 0.9792 ± 0.0010 | 202.901 ± 0.914 | 25.299 ± 0.015 | 0.91173 ± 0.00022 | 2.934 ± 0.001 |
| ECHO-H | 0.9748 ± 0.0010 | ± 0.718 | 24.496 ± 0.014 | 0.846 ± 0.008 | 2.760 ± 0.006 |
HDR Environment
| Method | SR ↑ | ↓ | ↓ | ↓ | ↓ |
|---|---|---|---|---|---|
| Vanilla | 0.8632 ± 0.0008 | 230.001 ± 0.833 | 31.541 ± 0.018 | 1.088 ± 0.003 | 3.701 ± 0.004 |
| Any2Track | 0.8455 ± 0.0009 | 227.549 ± 0.999 | 31.939 ± 0.012 | 1.087 ± 0.003 | 3.709 ± 0.003 |
| RMA-Teacher | ± 0.0021 | 216.826 ± 1.462 | 30.063 ± 0.017 | ± 0.002 | ± 0.002 |
| RMA-Student | 0.8700 ± 0.0009 | 224.568 ± 0.894 | 30.542 ± 0.014 | 1.065 ± 0.002 | 3.648 ± 0.001 |
| ECHO-PO | 0.8718 ± 0.0080 | ± 1.789 | 30.466 ± 0.116 | 1.056 ± 0.005 | 3.612 ± 0.025 |
| ECHO-SO | 0.8653 ± 0.0063 | 221.908 ± 2.865 | ± 0.012 | 1.067 ± 0.004 | 3.648 ± 0.017 |
| ECHO-H | 0.8922 ± 0.0016 | 210.351 ± 1.017 | 30.463 ± 0.008 | 1.039 ± 0.003 | 3.530 ± 0.003 |
Table 5
Motion-tracking evaluation on AMASS-2000-Random. Results are reported as mean sample standard deviation over three runs with action-noise standard deviation . Bold and underlined values indicate the best and second-best means within each environment, respectively.
Nominal Environment
| Method | SR ↑ | ↓ | ↓ | ↓ | ↓ |
|---|---|---|---|---|---|
| SONIC v1.1 | 0.9633 ± 0.0003 | 316.651 ± 1.649 | 30.411 ± 0.015 | 1.135 ± 0.002 | 3.770 ± 0.006 |
| HEFT | 0.9785 ± 0.0009 | 261.573 ± 2.352 | 28.524 ± 0.009 | 1.062 ± 0.001 | 3.103 ± 0.004 |
| SCALEBFM-XL | 0.9838 ± 0.0003 | 178.604 ± 0.622 | 31.179 ± 0.011 | 0.979 ± 0.005 | 2.753 ± 0.009 |
| BFM-ZERO | 0.6642 ± 0.0013 | 710.468 ± 7.304 | 88.765 ± 0.231 | 1.397 ± 0.002 | 7.179 ± 0.027 |
| P-BFM | 0.8485 ± 0.0031 | 631.253 ± 3.676 | 89.705 ± 0.323 | 1.620 ± 0.004 | 5.944 ± 0.006 |
| ECHO-PRE | ± 0.0012 | 207.995 ± 0.970 | ± 0.007 | ± 0.003 | 3.036 ± 0.006 |
| ECHO-N | 0.9827 ± 0.0003 | ± 0.289 | 25.577 ± 0.024 | 0.917 ± 0.003 | ± 0.007 |
NDR Environment
| Method | SR ↑ | ↓ | ↓ | ↓ | ↓ |
|---|---|---|---|---|---|
| SONIC v1.1 | 0.9545 ± 0.0022 | 343.709 ± 6.533 | 31.410 ± 0.056 | 1.196 ± 0.002 | 3.965 ± 0.004 |
| HEFT | 0.9738 ± 0.0008 | 326.887 ± 10.853 | 29.882 ± 0.157 | 1.123 ± 0.007 | 3.351 ± 0.032 |
| SCALEBFM-XL | ± 0.0029 | 245.105 ± 4.949 | 32.691 ± 0.225 | 1.061 ± 0.004 | 3.062 ± 0.008 |
| BFM-ZERO | 0.6503 ± 0.0029 | 818.227 ± 18.515 | 96.278 ± 0.426 | 1.512 ± 0.009 | 7.680 ± 0.057 |
| P-BFM | 0.8078 ± 0.0051 | 651.868 ± 12.303 | 93.229 ± 0.754 | 1.778 ± 0.014 | 6.233 ± 0.083 |
| ECHO-PRE | 0.9775 ± 0.0023 | ± 2.195 | ± 0.067 | ± 0.004 | 3.149 ± 0.012 |
| ECHO-N | 0.9803 ± 0.0012 | 214.227 ± 0.462 | 27.232 ± 0.090 | 0.943 ± 0.003 | ± 0.006 |
Response to Reviewer SYQw for question Q1
Thank you very much for carefully reviewing our paper and raising insightful questions that help us improve our work.
Capability of CRA has been shown by ECHO-H. In fact, we have already demonstrated some capabilities of CRA in our original demos, such as precise manipulation under varying payloads. Additionally, Video S2 illustrates the process of adapting to payload changes. However, these demonstrations mainly focus on the capabilities of ECHO-H trained under HDR, without explicitly highlighting the differences between ECHO-PRE and ECHO-N, both of which are trained under NDR. We will further show the capability of ECHO-N comparing with ECHO-PRE.
Interactive simulation demonstration. CRA is not merely a model, but rather a method that enables BFMs to identify and adapt to different environments, thereby enhancing their adaptability to diverse environments, especially extremely challenging ones. ECHO-N exhibits greater robustness in certain extreme environments, for example, on a very smooth surface, ECHO-PRE fails to maintain balance, whereas ECHO-N remains stable. Furthermore, we observed that providing ECHO-N with dismatched context degrades its tracking performance and may even lead to failure. To demonstrate these capabilities, we developed an interactive web demo, available at ECHO_N_vs_ECHO_PRE.
Real world results. To further highlight the differences between ECHO-PRE and ECHO-N, we also conducted additional real-world experiments, as illustrated in Video S1. On a smooth floor, ECHO-PRE fails to maintain balance, whereas ECHO-N succeeds. Specifically, we teleoperated the G1 robot to walk on the smooth floor to collect matched context and on a rough surface (a floor mat) to collect mismatched context.