RoboACE: Adversarial Co-Evolution for Robot Self-Improvement

Delin Zhao, Zhengrong Yue, Junlin He, Yali WangarXiv 2026 · arXiv

TL;DR

RoboACE is a framework for robot self-improvement that jointly evolves how a robot performs tasks and how it chooses what to practice. By using an adversarial co-evolution loop, the system systematically uncovers execution weaknesses and develops reusable software revisions to fix them, achieving state-of-the-art performance on complex manipulation benchmarks without fine-tuning underlying model weights.

What Problem Does This Paper Address?

Most current robot self-improvement systems focus on refining execution—improving how a robot handles a specific set of tasks based on feedback. However, if the source of practice tasks remains fixed, the robot eventually stops finding new weaknesses.

As a robot's capabilities grow, the "useful" practice it needs changes. A skill that works in isolation might fail when composed with others or placed in a cluttered environment. To keep improving, the robot needs a curriculum that adapts to its current skill level. The authors argue that the procedure for selecting practice must be as revisable as the execution logic itself.

Methods

The core of the paper is RoboACE (Robot Adversarial Co-Evolution). Instead of training the weights of large foundation models, RoboACE evolves the "harnesses" around them—the prompts, code-based skills, tools, and agent-readable memory that organize robot behavior.

The framework operates in an alternating, two-phase adversarial loop:

  1. Phase 1: Curriculum-Harness Evolution: A curriculum agent uses evidence from previous rounds to propose informative challenges (new tasks or world configurations) designed to expose gaps in the robot's current capabilities. An optimizer refines the logic the agent uses to select these tasks.
  2. Phase 2: Execution-Harness Evolution: The execution agent attempts the newly admitted challenges. When it fails, an execution optimizer diagnoses the failure and installs reusable revisions—such as better skill descriptions, improved recovery strategies, or refined controller handoff logic.
The alternating co-evolution loop between curriculum and execution harnesses.

To ensure stability, RoboACE uses an Admission Gate to verify that proposed challenges are physically valid and reachable before they are used for evaluation. It also features a Strategic Advisor that reviews long-term history to suggest broad research directions when local "whack-a-mole" fixes stall.

Datasets

The authors evaluated RoboACE within the LIBERO simulation ecosystem, a benchmark for lifelong robot learning. They used four distinct panels, each containing 200 conditions:

  • LIBERO: 40 tasks with reserved initializations (transfer to reserved conditions).
  • LIBERO-Plus: Tests robustness across seven perturbation categories.
  • LIBERO-PRO: Features four complex perturbation dimensions.
  • LIBERO-X: Five levels of increasing challenge, including language-paired conditions.

Crucially, the system only evolved using LIBERO conditions; the other three panels served as "held-out" transfer tests to measure the generalizability of the evolved harnesses.

Key Insights and Ideas

  • Software as the Evolvable Material: By evolving the "harness" (prompts and code) rather than model weights, the system remains interpretable and computationally efficient. Revisions are human-readable and can be audited.
  • Adversarial Synergy: The curriculum doesn't just get "harder"; it targets specific failures. For example, if the robot fails at opening drawers, the curriculum might generate tasks that require opening a drawer while holding an object, forcing the execution harness to develop better grounding and recovery logic.
  • The Importance of "Admission": You can't just generate random hard tasks. The "Admission Gate" ensures that the curriculum provides a "fair" and solvable challenge, which is essential for generating a clean signal for the execution optimizer.
  • Strategic vs. Local Optimization: Local fixes often address symptoms (e.g., "grip harder"). The Strategic Advisor identifies root causes (e.g., "when should we switch from general VLA control to a specific code-based skill?"), leading to more robust architectural changes.

Key Results

The final RoboACE deployment outperformed all evaluated baselines, including state-of-the-art controllers like $\pi_{0.5}$ and agent systems like VoLo.

  • Overall Gains: Compared to its initial state, the evolved RoboACE improved success rates by 35.5 percentage points on LIBERO and 30.0 points on LIBERO-Plus.
  • Superiority over Raw Controllers: On the challenging LIBERO-PRO panel, RoboACE reached a 74.0% success rate, exceeding the direct execution of its underlying $\pi_{0.5}$ controller by 13.5 points.
  • Hardest Tasks: On LIBERO-X, the most difficult panel, RoboACE achieved 30.5% success, nearly tripling the performance of the standalone $\pi_{0.5}$ (11.5%).
Success rates across the four evaluation panels compared to various baselines.

Limitations and Future Work

  • Simulation vs. Reality: The study was conducted entirely in simulation. While LIBERO is a high-fidelity benchmark, the "sim-to-real" gap remains a hurdle for physical deployment.
  • Computational Overhead: The co-evolutionary process requires significant simulation time and many foundation model calls to generate and test revisions.
  • Foundation Model Ceiling: The system's ultimate performance is still bounded by the reasoning and perception capabilities of the underlying frozen models (e.g., GPT-4).

The authors suggest that the next logical step is establishing this behavior across broader environments and moving toward physical robot campaigns.

Summary updated

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.
BibTeX
@misc{zhao2026roboace,
  author = {Delin Zhao and Zhengrong Yue and Junlin He and Yali Wang},
  title = {{RoboACE: Adversarial Co-Evolution for Robot Self-Improvement}},
  year = {2026},
  eprint = {2611.13034},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13034},
}
Figure