# Beyond Policy

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Vinal Gadhiya](https://cerebralvalley.ai/u/vinal)
- **GitHub:** https://github.com/vinal-gadhiya/ReActuate
- **Demo video:** https://drive.google.com/file/d/1jnaXZgjSnHIuhsQUGGFdE9rZEnuELey6/view
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3814

ReActuate: Reasoning and Recovery for Frozen Robot Policies

ReActuate is a closed-loop agentic robotics framework that uses reasoning, visual feedback, and adaptive planning to guide pretrained robot policies through long-horizon manipulation tasks without retraining the underlying models.

The goal is to explore whether reasoning and visual verification can improve the reliability of frozen Vision-Language-Action (VLA) policies and World Action Models (WAMs), particularly when tasks require multiple steps, object tracking, and recovery from execution failures.

ReActuate integrates Meta Muse Spark 1.3 Contributor for task planning, reasoning, and verification, with optional Meta SAM 3.1 for visual grounding. It operates in RoboCasa + MuJoCo, using frozen robot policies including GR00T N1.5, PatchWAM, and Awomo-0.5I, with model inference running on the NVIDIA Jetson AGX Thor.

Rather than generating a plan once and executing it blindly, ReActuate continuously observes the environment through multiple camera views, breaks tasks into smaller subtasks, and guides the robot policy using language instructions. After execution bursts, Muse reviews updated observations, verifies subtask completion using visual evidence, and adjusts the plan when something goes wrong.

In qualitative experiments on RoboCasa's Composite-Unseen tasks, ReActuate demonstrated several promising behaviors:

Adaptive planning: Maintaining task context and object awareness across execution steps.

Failure detection and recovery: Identifying unsuccessful placements and initiating corrective actions.

Evidence-based verification: Helping prevent premature transitions to the next subtask.

Policy-dependent outcomes: Demonstrating that better reasoning does not always translate into successful execution when the underlying policy cannot reliably follow instructions.

For example, in a pastry manipulation task, Awomo with Muse correctly selected and placed the requested croissant, while the baseline moved an additional unwanted object. In a strawberry placement task, Muse detected when the strawberry fell off the pancake and initiated a retry instead of treating the placement as complete.

These experiments are qualitative, and a full benchmark evaluation has not yet been completed. ReActuate therefore does not claim a measured improvement in aggregate task success rates.

The central idea is simple: robot policies shouldn't just execute instructions. They should be able to receive feedback, recognize failures, and adapt their behavior without requiring retraining.

## More from Meta Global AI Developer Hackathon

- [Kurdolingo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3811)
- [Harvest Relay](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3812)
- [Settleye](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3813)
- [Ultimate Truth](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3815)
- [DemoBot](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3816)
- [HOLD](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3817)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3814. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt