PI0.5 Reasoning — Step 76,200
This repository contains a LeRobot PI0.5 policy fine-tuned for language-conditioned manipulation and reasoning experiments with an SO-101 robot. It is the checkpoint saved at training step 76,200, corresponding to 12 data-equivalent epochs of the 1,200-episode training dataset.
The checkpoint is released as an artifact accompanying an ICRA paper submission. Quantitative paper results and citation information will be added when available.
Model summary
| Item | Value |
|---|---|
| Base checkpoint | lerobot/pi05_base |
| Framework | LeRobot 0.6.0 (30da8e687a6dfc617fcd94afc367ac7071c376ce) |
| Checkpoint | Step 76,200 / epoch 12 |
| Robot | SO-101, single-arm manipulation |
| Training objective | Supervised imitation learning with PI0.5 action flow matching |
| Cameras used | observation.images.middle, observation.images.wrist |
| Robot state | Six absolute joint positions |
| Actions | Six absolute joint-position commands |
| Action chunk | 50 steps |
| Action steps returned | 50 |
| Precision | bfloat16 |
Only the middle and wrist camera streams were provided to the policy. Other camera streams present in the dataset were not included in the model input features.
Training data
The model was trained on justintiensmith/VLA_Benchmark_Prompted_1200, containing 1,200 demonstrations across six manipulation/reasoning task families.
Dataset revision disclosure
This run was launched without an explicit dataset revision in its saved training configuration (dataset.revision: null) and therefore used the Hub main snapshot available at training time. The experiment records associate the run with the pre-trim snapshot 96a71744568016d553f2304c94ae393b1b28a148, containing 609,565 frames.
That snapshot predates a later correction which removed the first 0.4 seconds (12 frames at 30 FPS) of episode 819 because of an incorrect sensor reading. The corrected dataset revision is f1196a7fa8609a8aa44ed264b045017f55ab0a51, containing 609,553 frames. This PI0.5 checkpoint was not retrained after that correction. The VLA0 and MolmoACT2 comparison runs use the corrected dataset revision.
This difference is disclosed for reproducibility and should be considered when making controlled comparisons among the models.
Training configuration
| Parameter | Value |
|---|---|
| Training steps completed | 76,200 |
| Steps per epoch | 6,350 |
| Per-GPU batch size | 24 |
| Number of training processes / GPUs | 4 |
| Global batch size | 96 |
| Samples processed | 7,315,200 |
| Hardware | 4 × NVIDIA GH200 120 GB |
| DataLoader workers | 8 per process |
| Seed | 1000 |
| Image augmentation | Disabled |
| Video backend | PyAV |
| Held-out training evaluation split | None (eval_split=0.0) |
| Gradient checkpointing | Enabled |
| Vision encoder frozen | No |
| Expert-only training | No; the full trainable policy was fine-tuned |
| Model compilation | Disabled |
Normalization
{
"ACTION": "QUANTILES",
"STATE": "QUANTILES",
"VISUAL": "IDENTITY"
}
Optimizer and learning-rate schedule
| Parameter | Value |
|---|---|
| Optimizer | AdamW |
| Peak learning rate | 2.5e-5 |
| Betas | (0.9, 0.95) |
| Epsilon | 1e-8 |
| Weight decay | 0.01 |
| Gradient clipping norm | 1.0 |
| Warm-up steps | 1,000 |
| Cosine-decay steps | 30,000 |
| Final learning rate | 2.5e-6 |
The original training target was 101,600 steps (16 epochs). This repository intentionally preserves the step-76,200 checkpoint used for the reported epoch-12 evaluation.
Loading the policy
Install LeRobot 0.6.0 and load the model with:
from lerobot.policies.pi05.modeling_pi05 import PI05Policy
policy = PI05Policy.from_pretrained(
"justintiensmith/pi05_Reasoning_Step_076200"
)
For evaluation or robot deployment, use the preprocessor and postprocessor files shipped in this repository together with the model checkpoint. They contain the normalization statistics and transformations used during training.
Evaluation status
Preliminary real-robot rollouts indicated useful task behavior at the epoch-10 and epoch-12 checkpoints. This statement is qualitative: no quantitative success rate is claimed in this model card yet. Controlled per-task results, evaluation protocol, and uncertainty estimates will be reported with the associated paper.
Intended use
This checkpoint is intended for:
- reproducible robotics research;
- comparison of vision-language-action policies on the associated benchmark;
- real-robot evaluation with a compatible SO-101 observation/action interface;
- analysis of language-conditioned manipulation and reasoning behavior.
Limitations and safety
- The policy was trained on a limited set of robot demonstrations and may not generalize beyond the recorded camera views, workspace, objects, prompts, or robot calibration.
- The training data contain correlated frames from demonstrations; the number of frames is not the number of independent examples.
- No held-out loss split was used during training. Real-robot task success, rather than training loss, should be used for model selection.
- This checkpoint used the pre-trim dataset snapshot described above.
- Outputs can be unsafe or unpredictable under distribution shift. Use physical safeguards, conservative motion limits, emergency-stop procedures, and human supervision when operating real hardware.
Citation
Citation details will be added when the associated ICRA submission is publicly available. Until then, please cite this model repository and the accompanying dataset with their pinned revisions.