MiniRDT-SO101

A manipulation benchmark can look successful while being physically invalid. An audit found the original scene allowed table penetration, invalidating its earlier grasp results.

• A physics-v2 MuJoCo task with table contact, calibrated contacts, a side-pinch expert, and explicit success checks.
• TinyRDT: a 4-layer Transformer conditioned on vision, proprioception and diffusion time, predicting 16 joint-target actions.
• Data collection, closed-loop diagnostics, generalization probes and regression tests alongside the policy code.

The policy runs receding-horizon control: observe, sample an action chunk, execute a prefix, and observe again. The evaluation distinguishes memorized scenes from held-out positions and logs contact validity rather than relying on a single visual success signal.

The project keeps invalid physics-v1 results in the history rather than presenting them as wins. Offline loss was not enough to predict closed-loop skill, and spatial coverage mattered more than raw demonstration count in the documented experiments.