AnchorVLA: Anchored Diffusion for Efficient End-to-End Mobile Manipulation

A team of researchers, including Jia Syuen Lim, Zhizhen Zhang, and others, has introduced AnchorVLA, a novel AI architecture designed to solve a core problem in mobile robotics: balancing the need for multiple possible action plans with the requirement for fast, reactive control. Traditional diffusion models are excellent at generating diverse solutions (like different ways to grasp a bottle in clutter) but are computationally expensive because they require many iterative denoising steps. AnchorVLA's key innovation is 'anchored diffusion,' which starts the denoising process near a plausible solution, allowing it to use a much shorter, truncated schedule. This retains the model's crucial multimodality while making inference fast enough for real-time, closed-loop robot control.

Beyond speed, the system tackles another major issue: action chunking. To amortize computation, robots often plan a sequence of actions (a chunk) and then execute it open-loop, which can lead to accumulating errors and drift. AnchorVLA incorporates a lightweight residual correction module that runs at a high frequency during rollout, making small per-step adjustments to the planned trajectory. This test-time self-correction mechanism allows the robot to stay on course even when faced with unexpected disturbances or shifts from its training data. The result, demonstrated across diverse mobile manipulation tasks, is a policy that is both more robust and more efficient, achieving higher success rates with lower-latency inference. The source code has been made publicly available, facilitating further research and application.