Upload Piper adaptive B-spline README.md

models/piper-pick-white-block-25hz-qwenpi-v3-adaptive-left-exclude-gripper-native-controls13-gb32-50k-20260905/README.md

@@ -0,0 +1,70 @@
+# Piper QwenPI_v3 25 Hz Adaptive-Left B-spline — GB32 50k
+
+This checkpoint trains `QwenPI_v3_AdaptiveBSpline` on all 52 episodes of
+`20260818_piper_pick_white_block_25hz`. It uses the native flow-matching loss
+only; no L1 action or reconstruction loss was added.
+
+## Action representation
+
+- Six arm joints plus gripper (`7` physical action dimensions)
+- Continuous degree-3 `adaptive_left` B-spline
+- Nominal knot span: 2 action steps; nominal chunk: 20 actions at 25 Hz
+- 13 rows per prediction, each `[7 control values, duration_steps]`
+- Gripper dimension 6 is excluded from knot insertion and fit once after the
+  final arm-derived knot placement
+- Native execution length is `sum(first 10 predicted durations)`; plans are
+  never resampled to a fixed horizon
+- Observed target native horizon: 10–48 steps; deployment safety cap: 64
+
+All 17,141 encoded observations passed exact control/duration interleaving,
+knot reconstruction, and normalized deployment-decoding checks. At the same
+global control budget as uniform-left, adaptive-left reduces episode-wide arm
+representation RMSE by `31.69%`. On the common guaranteed 10-step prefix, it
+reduces arm representation RMSE by `20.17%`.
+
+## Training
+
+- 50,000 optimizer steps; checkpoints every 10,000 steps
+- Global batch 32 (8 per GPU on four GPUs, no accumulation)
+- Four NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs
+- BF16 with DeepSpeed ZeRO-2
+- Frozen Qwen3-VL-4B-Instruct backbone
+- Native flow-matching velocity MSE only; no L1 loss
+- Wall time: 7 hours 48 minutes 11 seconds
+- W&B: https://wandb.ai/401910710-university-of-california-berkeley/starVLA_piper/runs/99wnbenv
+
+## Full open-loop evaluation
+
+Every one of the 17,141 recorded observations was evaluated independently.
+Adaptive plans were decoded at native 25 Hz and masked at their predicted
+duration sum. This is a training-set open-loop evaluation, not held-out
+generalization.
+
+- Predicted native horizon: 10–49 steps, mean `19.96`
+- Target native horizon: 10–48 steps, mean `20.40`
+- Horizon MAE: `1.33` action steps
+- Physical-action RMSE on the predicted native horizon: `0.02307`
+- Arm RMSE: `0.01570`; continuous gripper RMSE: `0.04739`
+- First-action RMSE: `0.01255`
+- Binary gripper accuracy after the deployment constraint: `98.37%`
+- MSE reduction versus repeat-current-action baseline: `94.15%`
+
+The common 10-step comparison against the uniform-left 50k checkpoint shows:
+
+- Representation arm RMSE: adaptive `0.0004592`, uniform `0.0005753`
+  (`20.17%` lower for adaptive)
+- Policy arm RMSE: adaptive `0.01473`, uniform `0.01380`
+  (`6.72%` higher for adaptive)
+- Overall policy RMSE: adaptive `0.01995`, uniform `0.01700`
+  (`17.36%` higher for adaptive)
+
+Thus the adaptive representation advantage is verified, but it did not
+translate into better 50k policy prediction under this native flow-matching
+training setup. Gripper is reported separately because it deliberately does
+not place knots.
+
+## Artifact
+
+Deployable checkpoint: `checkpoints/steps_50000_pytorch_model.pt`
+
+SHA-256: `23656dc0823fdbe8b1f285b83d451fdbb930f6f1789c1f39d4ab98729e64f9f6`