AlexWortega/tinyvla-microduck-stage25

TinyVLA MicroDuck Stage 2.5

Experimental Stage 2.5 checkpoint with a 100-action (two-second at 50 Hz) native joint-offset horizon. It continues the MicroDuck TinyVLA training with ordinary task, composed multi-stage, recovery/balance, and transition-context streams.

• input: corrected 256×256 physical head_camera RGB, normalized 61D state, and one fixed high-level instruction;
• output: 100×14 joint-position offsets in radians;
• embodiment id: 24;
• runtime: execute only five actions, capture a new 10 Hz observation, replan;
• recovery: autonomous from the state/vision context without substituting a stand-up language prompt.

See SERVING.md for the HTTP payload, state layout, environment, and launch command.

ONNX CPU serving

onnx/ contains the complete Torch-free serving export used by the public MicroDuck simulator Space. The vision encoder, cacheable semantic path, and 10-step flow-matching action sampler are separate fixed-shape ONNX graphs. onnx/metadata.json records the source checkpoint SHA256, graph contracts, checksums, and PyTorch/ONNX parity results. The original safetensors checksum is 8a603a3a0c19961e74e8cb1f49fe30bb6a0712f6d5e5fba8429bdbb69a4826a8.

Interim evaluation

The bundled task-only offline evaluation covers 90 samples and reports normalized RMSE 0.8555 versus the zero-action baseline 0.9191. Dedicated recovery buckets are intentionally omitted from that number. Closed-loop Stage 2.5 evaluation is still in progress, so these weights must be treated as an interim research checkpoint.

Terms and safety

This checkpoint is simulation-only and is not safety-qualified for physical hardware. Keep joint/current limits and an independent emergency stop enabled. The upstream MicroDuck 3D assets used in the rendered training data are marked CC BY-NC-SA; the repository uses license: other because model-weight and source-asset licensing are distinct. Review the upstream TinyVLA and MicroDuck terms before redistribution.