AFIL Uses Failure Trajectories to Boost Robot Manipulation Success Rates

Vision-language-action (VLA) models are a promising paradigm for robotic manipulation, but they typically rely on behavioral cloning from successful demonstrations only. This leaves them brittle: small execution errors compound into unrecoverable failures. To address this, researchers from GE HealthCare and University of Illinois at Urbana-Champaign introduce Adaptive Failure-Informed Learning (AFIL). The key insight is to use failure trajectories as adaptive negative guidance during training. AFIL leverages a pretrained VLA model to generate failure rollouts online, eliminating the need for handcrafted failure modes or human-in-the-loop recovery. It then trains Dual Action Generators (DAGs) — one for successful actions, one for failed behaviors — sharing a common vision-language backbone. This design keeps parameter overhead low while enabling failure-aware learning.

During sampling, the failure generator adaptively steers action generation away from failure-prone regions and toward success modes. The steering strength is determined by the per-diffusion-step distance between success and failure distributions, making guidance dynamic and context-aware. Experiments cover both short- and long-horizon manipulation tasks, in-domain and out-of-domain settings. Results show AFIL consistently improves task success rates and robustness over existing VLA baselines without requiring additional demonstrations or human feedback. The method is model-agnostic, working with diffusion- and flow-based policies. This work points toward a future where robots learn from errors autonomously, a critical step for real-world deployment.