Skip to content

RealityPatch

RealityPatch is an AI-powered platform that diagnoses and repairs robot control failures in physics simulations.

Code Video Cerebral Valley

Core trick

Agents read simulation video and sensor logs, patch controller code, and rerun scenarios to score mission completion.

Tags

robotics/hardwareagents

Description

RealityPatch is an AI-powered platform that diagnoses and repairs robot control failures in physics simulations. AI agents inspect video and sensor data, modify controllers, and rerun tasks across robot dogs, drones, warehouse trolleys, and cars. An interactive dashboard compares their changes and measures whether each machine successfully completes its mission.
RandyLimmy/Astra2026Hackathon★ 00 forksPythonCreated Sep 10, 2026 · last push Sep 11, 2026

README excerpt

# RealityPatch

We lost lol. Maybe we were too research-oriented.

Anyways, here is our project. 

The current testing scope is **these four complete control tasks**:

| Machine | Scenario ID | Required successful outcome |
| --- | --- | --- |
| Robot dog | `quadruped_gait_failure` | Follow the requested speed transition along the walking strip while staying upright. |
| Drone | `drone_delivery_imbalance` | Deliver the parcel from A to B, release it on the pad, return unloaded, and land at A. |
| Warehouse trolley | `warehouse_curve_demo` | Complete the marked 90-degree bend and exit with the cargo aboard and no ground impact. |
| Car braking | `car_auto_brake_failure` | Stop in the marked zone before the wall without impact, using the fixed 22 m/s approach. |

The warehouse scene is a differential-drive trolley. Older fault-maintenance
presets remain in the repository as historical workflows, separate from the
current controller tasks and batch.

The car task declares a 22 m/s approach so a controller-only braking repair is
physically feasible. Its original preview and all model trials use this same
speed and the same four preparation cycles. The retained legacy preset's
25 m/s approach is unchanged.

Astra and GPT-5.6 Sol can inspect recorded RGB images, public sensor measurements,
the original controller and its permitted settings; install controller changes;
and run the **same complete task** again. The physical world, initial conditions,
mission and deadline stay fixed.

Team