Pi_atlas_RECAP
RECAP-style value functions for the YOR bimanual robot, trained on adityx23/Pi-Atlas-Full. One frame in, no history.
Each model is a PaliGemma backbone, initialized from openpi's pi05_base, plus a learned value token and a
201-bin value head. Training is cross-entropy on the binned target, and the prediction is the expectation over the
bins. The prompt is <task>. Subtask: <subtask>, and the input includes the 21-dim state. During training, 5% of
samples get a wrong subtask from the same trajectory, with target β1.
All models: 10k steps, batch 64, seed 0, 2 GPUs (FSDP).
Code: HannibalofBarca/pi_atlas_value_functions, an openpi fork (src/openpi/models/pi_value.py, configs in src/openpi/training/config.py).
Models
Each folder is an openpi/Orbax checkpoint: params/, the norm-stats assets/ and _CHECKPOINT_METADATA. The
optimizer state is left out.
| Folder | Config | Target | Step |
|---|---|---|---|
recap_full_terminal/ |
pi05_recap_full_terminal |
Terminal: β(steps to episode end) / max length, β1 on failure | 9999 |
recap_full_dense/ |
pi05_recap_full_dense |
Dense progress over the cycle lap: β((Nβ1βk) + (Tβ1βt)/(Tβ1)) / N, β1 on failure | 4000 (stopped early) |
recap_term_st/ |
pi05_recap_term_st |
Per-subtask terminal: β(Tβ1βt)/(Tβ1) within each subtask, β1 on failure | 9999 |
recap_term_st_h30/ |
pi05_recap_term_st_h30 |
As term_st, plus a second head for the value 30 frames ahead |
9999 |
recap_term_st_h30_bellman/ |
pi05_recap_term_st_h30_bellman |
As h30, plus Bellman agreement (Ξ» = 1) between the +30 head at t and the stop-gradient current head at t + 30, within a subtask |
9999 |
recap_term_st_critic/ |
pi05_recap_term_st_critic |
Q(s, a): term_st target, with the executed 30-step action chunk added to the prompt as FAST tokens |
9999 |
Holdout results
There are two holdout sets. "Old" is 28 episodes (10 red chili, 18 eggplant laps) and "New" is 126 episodes across all tasks. Each model is scored against its own target. Terminal, dense and per-subtask targets differ, so numbers are comparable only between models that share a target.
| Model | Old MAE | Old Pearson | Old Spearman | New MAE | New Pearson | New Spearman |
|---|---|---|---|---|---|---|
| recap_full_terminal | 0.070 | 0.734 | 0.773 | 0.138 | 0.607 | 0.744 |
| recap_full_dense | 0.065 | 0.782 | 0.783 | 0.047 | 0.923 | 0.914 |
| recap_term_st | 0.067 | 0.818 | 0.818 | 0.115 | 0.787 | 0.779 |
| recap_term_st_h30 | 0.073 | 0.771 | 0.773 | 0.121 | 0.766 | 0.751 |
| recap_term_st_h30_bellman | 0.065 | 0.806 | 0.806 | 0.115 | 0.785 | 0.778 |
| recap_term_st_critic | 0.076 | 0.790 | 0.789 | 0.133 | 0.731 | 0.721 |
Loading
from openpi.training import config as _config
from openpi.models import model as _model
cfg = _config.get_config("pi05_recap_term_st")
params = _model.restore_params("recap_term_st/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.