Agentic-VLA boosts robot adaptation with 2.4x faster learning

Agentic-VLA tackles two core weaknesses of current Vision-Language-Action (VLA) models: poor generalization to new environments and the need for massive demonstration datasets. The framework, developed by Ruofan Jin and Zaixi Zhang, introduces three key mechanisms that let robots continuously learn during deployment. Adaptive Reward Synthesis dynamically generates reward functions based on the VLA's current skill level and task complexity, breaking complex tasks into learnable sub-goals for curriculum learning. Language-Guided Exploration uses a critic model to provide structured guidance for systematic exploration rather than random trial-and-error. Experience Memory stores and retrieves policy weights from similar past tasks to warm-start adaptation, drastically reducing retraining time.

Evaluated on the LIBERO benchmark, Agentic-VLA delivers substantial gains: +12.3% on long-horizon tasks, +28.5% in 1-shot learning scenarios, and a jump from 0% to 31.2% in cross-task transfer zero-shot without any task-specific demonstrations. The framework also converges 2.4x faster compared to existing online adaptation methods. Beyond LIBERO, it retains its advantage on the dual-arm RoboTwin 2.0 benchmark, including under its randomized Hard setting. These results position Agentic-VLA as a significant step toward truly adaptive robotic systems that can learn and improve continuously in real-world environments.