Pi_atlas_KI
pi0.5 vision-language-action models for the YOR bimanual robot that predict a value alongside the actions. Trained on adityx23/Pi-Atlas-Full.
The backbone (initialized from openpi's pi05_base) outputs three things in order:
- The subtask as text (
Subtask: <subtask>;). - A value slot: a learned value token with a 201-bin value head. Its target is RECAP terminal per subtask episode: β(steps to end) / max length, β1 on failure.
- The action chunk as FAST tokens.
A flow-matching action expert, which attends to the images, the prompt and the subtask, produces the continuous actions. Action tokens don't attend to the value slot. During training, 5% of samples get a wrong subtask, which masks the subtask and action losses and sets the value target to β1.
The two models differ only in knowledge insulation (KI, Driess et al. 2025):
| Folder | Config | KI | Step |
|---|---|---|---|
ki/ |
pi05_atlas_KI_vla_ki |
On: the expert reads the backbone through a stop-gradient, so the flow loss doesn't train the backbone | 9999 |
noki/ |
pi05_atlas_KI_vla_noki |
Off: joint training | 9999 |
Both: 10k steps, batch 64, seed 0, 2 GPUs (FSDP). Each folder is an openpi/Orbax checkpoint: params/, the norm-stats
assets/ and _CHECKPOINT_METADATA. The optimizer state is left out.
Code: HannibalofBarca/pi_atlas_value_functions, an openpi fork (src/openpi/models/pi_atlas_KI_vla.py, configs in src/openpi/training/config.py).
Value-head holdout results
These are scored against the terminal target. "Old" is 28 episodes (10 red chili, 18 eggplant laps) and "New" is 126 episodes across all tasks. "Wrong-subtask AUROC" is measured on the new holdout. It is the AUROC for separating each frame's value under its true subtask from its value under a wrong subtask.
| Model | Old MAE | Old Pearson | Old Spearman | New MAE | New Pearson | New Spearman | Wrong-subtask AUROC |
|---|---|---|---|---|---|---|---|
| ki | 0.056 | 0.830 | 0.804 | 0.125 | 0.627 | 0.755 | 0.914 |
| noki | 0.061 | 0.789 | 0.794 | 0.154 | 0.583 | 0.735 | 0.896 |
The actions themselves have not been evaluated on the robot.
Loading
from openpi.training import config as _config
from openpi.models import model as _model
cfg = _config.get_config("pi05_atlas_KI_vla_ki")
params = _model.restore_params("ki/params")
model = cfg.model.load(params)
The configs expect the Pi-Atlas-Full LeRobot dataset and the norm stats in assets/. Point assets_dir and the
paths in the config at your local copies.
License
The models derive from openpi's pi05_base, which includes Gemma weights. Use is subject to the
Gemma Terms of Use and openpi's license.