ZebTaylor/panthera-ik-three-block-stack-30hz

Panthera IK three-block stacking

1,000 successful scripted IK demonstrations; 862,766 training frames at 30 Hz. This is the three-block Panthera MuJoCo task: red onto green, then blue onto red. The stored task prompt is stack the three colored cubes.

LeRobot v3.0 format, exported using LeRobot 0.4.4. Both 256×256 RGB cameras (observation.images.shoulder, observation.images.wrist) are embedded JPEGs in Parquet. No video decoder, raw-data conversion, or MuJoCo rendering is needed for training. Download data/** and meta/** to use LeRobotDataset.

Feature	Contract
observation.state	Six measured joint angles in radians, then current gripper opening command in metres
action	Six absolute joint position targets in radians, then absolute gripper opening command in metres
Timing	Observation at t predicts control at t + 1/30 s; actions already shifted once
Camera order	Shoulder, then wrist; preserve RGB and physical wrist roll

Do not apply another action shift, delta conversion, gripper sign inversion, or end-effector action transform. Short vectors are padded inside pi0.5, not in the stored data. These are joint-space labels even though an IK planner produced them.

The collector accepted 1,000 of 1,402 randomized attempts. All 1,000 accepted episodes passed a fresh physics replay of the resampled controls, including a released three-stack held for the final second. verification.json records the export audit: every action/state row matched its source, and 2,004 sampled camera images decoded successfully. These are demonstration checks, not learned-policy success rates. Starts are restricted to a reachable elevated region; rejected attempts introduce selection bias. Test learned policies on held-out seeds and report both matched starts and the broader teleoperation reset distribution.

meta/provenance.json retains source and renderer hashes. The training notebook uses a deterministic episode holdout, and computes normalization statistics on training episodes only. Dataset-wide statistics here are supplied for general LeRobot compatibility. Full-model pi0.5 fine-tuning must explicitly unfreeze the vision encoder and VLM and omit PEFT/LoRA.

Source project: https://github.com/zebulontaylor/RobotArmTraining Dataset: https://huggingface.co/datasets/FoxNerdSaysMoo/panthera-ik-three-block-stack-30hz

Full-model pi0.5 notebook and rollouts

Download the notebook for one RTX PRO 6000 with 96 GB VRAM. It includes fully unfrozen training, strict weight loading, a gradient/VRAM smoke test, resume support, embedded two-camera rollout videos, and fixed-seed task success evaluation with confidence intervals. The dataset/runtime and base model are pinned to immutable revisions.

The ~15 MB assets/pi05_runtime.zip contains the matching simulator meshes/scene, camera definitions, and training/evaluation helpers. The notebook retrieves and verifies these automatically. No local project checkout or data rendering is needed.

Validation and implementation notes distinguish local CPU/simulator checks from the full GPU training run, which the user executes.

Storage and W&B update

The notebook no longer requires 300 GiB free at startup. It reports storage, skips model/optimizer files during the smoke run, and keeps the latest two full training checkpoints (configurable to one). New checkpoints are saved completely before pruning older snapshots; the saver checks space against actual tensor sizes.

W&B is enabled by default for training/validation metrics. Rollout evaluations log success rates, confidence intervals, and optionally videos in grouped runs. API keys are entered securely; offline logging is supported. Model checkpoint artifact uploads stay disabled. Download the latest notebook for these updates.

Complete simulator assets and early Google Drive access

The runtime now contains both visual and collision meshes. It is compiled from a clean archive extraction before publication; both cameras were also rendered from that isolated extraction.

Google Drive backup is enabled by default in Colab. Its authorization/mount prompt comes before installation and downloads. Every completed training checkpoint is backed up with optimizer/RNG state to MyDrive/pi05_ik3/<RUN_NAME>/. Only the most recent complete backup is retained, after the new copy finishes successfully. Set RESUME_FROM_DRIVE=True with the same run name to restore the latest backup and verify its checksums. The smoke run saves no checkpoint.