A new reinforcement learning framework called HALO-WA patches a stubborn blind spot in robot manipulation: the final few millimeters of a task where most failures happen.
World-action models can generate long sequences of robot movements for general manipulation, but they break down on precision work - inserting a connector, aligning a part - because they have no way to correct for calibration drift or contact-force surprises at the last moment. HALO-WA adds a lightweight adapter on top of an existing world-action model, reading internal latent features from the model's own generation process and using them to refine the robot's action. The whole thing can be trained online - on the actual robot, in the actual environment - in 45 to 75 minutes per task. Across four real-world precision tasks, average success rates climbed from 26.4% to 87.1%, beating the next-best approach by 19.2 percentage points.
The result matters because "general-purpose" robot models have a credibility problem: they look impressive on broad benchmarks and fall apart on anything requiring tight tolerances. HALO-WA's approach of borrowing latents from the base model rather than training a separate corrective model keeps the adapter small and fast enough to actually run in deployment - a practical constraint most lab demos quietly ignore.
The code is public and the team ran supplementary tests in the RoboTwin simulator for reproducibility - a welcome step in a field where real-world results are notoriously hard to replicate. Whether the 45-to-75-minute training window holds outside the four tasks reported here is the obvious next question.