README.md

---
library_name: lerobot
tags: [robotics, lerobot, diffusion, putcab, train50, simulation]
datasets:
- Shiki42/PutCab-Sequential-Train50-V4
---
# PutCab sequential — diffusion — Train50 — E248

Full official LeRobot policy checkpoint after **10000 optimizer updates** from run `E248-R001`. Weights and normalization tensors are byte-identical to the independently verified training checkpoint. This repository is inference-only; optimizer/data-loader resume state is not included.

Dataset: [Shiki42/PutCab-Sequential-Train50-V4](https://huggingface.co/datasets/Shiki42/PutCab-Sequential-Train50-V4) at `c6022cf9164b34c7244dfede8ece4d20a999ab82`, 50 episodes/18143 frames. Training used the verified v3 conversion with original numerical columns and decoded RGB retained. Batch8, seed87431; precision `bf16 autocast, official optimizer/master parameters`. See `train_config.json` and `provenance.json` for the model-specific recipe and exact origins. The historical config field `use_amp` does not replace the recorded trainer mixed-precision setting.

Inputs: head, left-wrist and right-wrist RGB (320x240) plus14D commanded joint/gripper state. Outputs: absolute left6 joints+gripper,right6 joints+gripper. No delta conversion or implicit idle-arm masking is added by the upload. Prediction horizon16, executed queue8, denoising steps100; a queue length is not a denoising count.

## Loading

Use official LeRobot0.4.4/source `8fff0fde7c79f23a93d845d1a50e985de01f8b8a` with compatible Torch/Transformers. After `snapshot_download("Travor278/putcab-sequential-diffusion-train50-10k-e248", revision="<pinned repository revision>")`, import the included `load_policy.py` and call `load_policy(snapshot_dir, device="cpu")` (or `cuda`). It returns `(policy, preprocessor, postprocessor)`. Use the processors for normalization and denormalization; reset policy queues between episodes. Images must follow the checkpoint's named camera mapping and float RGB convention.

SmolVLA packages include pinned tokenizer/config assets under `vlm_assets/`; the loader resolves those locally. ACT loading skips a redundant ImageNet initialization download because its complete trained backbone is already included. Model weights are not converted or retrained. JSON changes only replace machine-local paths; these are recorded in provenance.

## Recorded evaluation

**0/100 successes**, 100 collision-marked episodes, on frozen PutCab-FixedRole100-v1 using the corrected per-scene100Hz protocol (`native-fixed100hz-v2`), RT4/depth4/OptiX, centered FOVY90 wrists, maximum700 model actions. Physical timestep is0.010s; each action uses the original variable-duration TOPP interpolation. This is not a claim of60ms action cadence or a new evaluation performed for publication.

Original incomplete infrastructure attempts were retained, and only missing seeds were recovered after full control input/action/initial-state qualification. Controls are excluded from the100-scene denominator. Old invalidated-protocol scores are not used here. Results remain **reported/audit pending**; low average training loss does not establish robust grasp/release behavior.

Evaluation adapter/diagnostic code: https://github.com/Shiki42/CTR/tree/travor/eval/robotwin/native-policy . Publication manifests give per-file SHA256; the source checkpoint passed CPU finite checks, and this portable package is separately reloaded offline before upload.