DVIDIA 路 PiPER learned feedback v0.2
A simulation research pilot: a learned controller paired with the frozen v0.1 motion primitive, trained on synthetic authored-expert labels and privileged simulator sensing. The matched blind control is separately trained without live feedback. Native servos, actuator limits and gravity compensation remain authored. This is not human-video learning, general slip recovery, hardware qualification, or proof of one-shot robot skill transfer.
| Variant | Strict successes / scheduled | Valid |
|---|---|---|
| feedback | 6 / 24 | 24 |
| blind | 8 / 24 | 24 |
| feedforward | 5 / 24 | 24 |
| teacher | 24 / 24 | 24 |
| open_jaw | 0 / 24 | 24 |
All scheduled outcomes are retained, including failures. The dataset contains original 1 ms native NPZ arrays, teacher demonstrations, every development candidate, the DAgger collection (if used), confirmation traces, sources and receipts. The static replay is a smaller recorded projection; it does not run physics or inference in the browser.
When training-parent.json is present, original teachers and failed experiments live unchanged under prior-studies/study01/ with their original protocol and source hashes. Preserving the old DAgger collection does not imply that the selected new model trains on those mixed-policy samples; consult its bound training receipt for the actual recipe.
Model 路 Complete evidence dataset 路 Recorded replay 路 Research paper 路 Source and local CLI
Limits
- Synthetic authored-expert jaw labels and privileged simulator sensing; not human video or hardware.
- Only the jaw is controlled by the new learned adapter; all six arm commands remain the frozen v0.1 primitive.
- Same rigid 40 g block, two trained destinations and narrow starting region as v0.1.
- Development starts were reused after five full-residual candidates each failed all 16 development cases.
- Interventions affect acquisition; airborne regrasp and general slip recovery are not established.
- Native servos, joint/jaw caps and gravity compensation remain authored.
- Timing excludes sensing, network transport and physical actuators; no hard real-time guarantee.
Original DVIDIA code and artifacts: MIT License. Upstream robot assets and rendering components retain their licenses; notices are included where applicable. Hashes establish artifact consistency, not independent authenticity.