RoboACE: Adversarial Co-Evolution for Robot Self-Improvement
Abstract
Robot agents can improve reusable skills, planning procedures, and recovery strategies from execution feedback, yet a fixed practice source leaves the search for new weaknesses unchanged. Useful practice depends on the agent's current capabilities, so the procedure that chooses practice must also be revisable. We present RoboACE (Robot Adversarial Co-Evolution), a framework that jointly evolves curriculum and execution harnesses around frozen foundation models. The curriculum harness uses accumulated capability evidence to propose informative challenges, while the execution harness responds with reusable revisions that expand execution capabilities. Alternating updates hold one side fixed as the other changes, and admitted challenges form fixed snapshots for comparing execution revisions. Starting from a compact set of LIBERO practice conditions, the resulting frozen deployment outperforms the evaluated controllers and agent systems across four LIBERO-family panels. On LIBERO-PRO, it reaches 74.0% success, exceeding the initial deployment by 15.0 percentage points and direct execution by its VLA controller by 13.5 points. Case studies document changes to evidence-guided practice selection, drawer-skill grounding, and controller intervention.
Figures & tables
| Panel | Coverage | Evaluation role |
|---|---|---|
| LIBERO | 40 tasks, five reserved initializations | Transfer to reserved conditions |
| LIBERO-Plus | Seven perturbation categories | Transfer beyond development sources |
| LIBERO-PRO | Four available perturbation dimensions | Transfer beyond development sources |
| LIBERO-X | Five challenge levels | Transfer beyond development sources |
| System | LIBERO | LIBERO-Plus | LIBERO-PRO | LIBERO-X |
| [ 1 ] | 96.9 | 92.0 | 60.5 | 11.5 |
| OpenVLA-OFT [ 12 ] | 96.8 | 75.5 | 51.0 | 2.5 |
| -FAST [ 24 ] | 85.5 | 61.6 | 49.0 | 4.5 |
| Cosmos Policy [ 14 ] | 98.5 | 88.5 | 52.5 | 2.5 |
| GR00T N1.7 [ 22 ] | 97.5 | 89.0 | 48.0 | 2.5 |
| RPent [ 34 ] | 92.0 | 84.0 | 32.5 | 12.5 |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Step | Independent work and revision | Fixed reference or output |
|---|---|---|
| 1 | Curriculum agent proposes practice; its optimizer tests reusable revisions. | Execution deployment and prior profile . |
| 2 | Validate proposals, retain or replace practice, and curate reference coverage. | Fixed comparison snapshot . |
| 3 | Evaluate the incoming execution deployment and compare reusable revisions. | Same tasks, initializations, settings, and budgets. |
| 4 | Adopt Candidate or retain Stable; consolidate permitted evidence. | Selected deployment and next profile . |
Robot Learning
Robot learning develops methods for robots to acquire skills and adapt from demonstrations, interaction, data, and physical models.
See all 340 papers