Research map / Vision-language and multimodal / Robot action and navigation
RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RePO-VLA, a recovery-driven VLA training framework (history-reset recovery initialization plus value-conditioned refinement with a progress-aware semantic value function) and FRBench, a recovery benchmark with error injection in RoboTwin simulation and real bimanual tasks.
Problem
VLA policies trained mainly on successful demonstrations are brittle under execution drift in long-horizon, contact-rich bimanual manipulation, and failed or recovery rollouts are typically discarded or imitated naively.
Approach
Assigns distinct roles to success, recovery, and failure trajectories: recovery segments are sliced with observation history reset to avoid causal confusion, a self-referential semantic progress value function provides dense labels (salvaging failure prefixes via reliability decay), and a value token conditioned on v=1.0 at deployment steers the pi0.5 policy back toward the success manifold.
Key results
- On FRBench-Sim injected-failure split, RePO-VLA Full averages 37.0 (clean) and 43.0 (randomized) vs pi0.5 at 15.0 and 15.4.
- On real-world adversarial tasks, Full (4x recovery data) averages 75% success vs 20% for pi0.5 (reported as 55-point gain).
- Phase I (w/o Fail) keeps nominal randomized average at 44.0 vs 33.6 for pi0.5, while pi0 drops to 12.9 under randomization.
Surprising findings
- Under real-world adversarial perturbation, the 1x Full model underperforms the Phase I-only variant (30 vs 37.5), which the authors attribute to data sparsity.
Limitations
- Relies on observing representative failure modes; out-of-taxonomy errors reduce zero-shot recovery.
- Real-world recovery data via teleoperation is costly.
- FRBench-Sim does not cover fluids or highly deformable objects.
Experimental setup
Datasets: RoboTwin FRBench-Sim
Benchmarks: FRBench-Sim FRBench-Real
Metrics: task success rate
Baselines: pi0 pi0.5 RDT GO-1 heuristic retry baseline
Builds on: pi0.5 V-JEPA RoboTwin
Closest papers by meaning
- Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.90)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.90)
- Human-assisted Robotic Policy Refinement via Action Preference Optimization (NeurIPS 2025 · similarity 0.89)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.