PranayTest/piper-left-clean-rad-v1-act-front-top-left
This is the selected offline imitation-learning checkpoint for the PiPER
left-arm plate pick-and-lift task. It was selected from parallel ACT and
SmolVLA ablations using evaluation action-prediction error and the requested
train-test divergence criterion.
Selected checkpoint
- Run:
act-front-top-left-chunk40
- Ablation label:
act-front-top-left
- Training step:
10000
- Train standardized MAE:
0.117780
- Eval standardized MAE:
0.173132
- Test standardized MAE:
0.191308
- Absolute train-test divergence:
0.073528
- Test joint MAE:
0.072642 radians
- Test gripper MAE:
0.243458
- W&B project: https://wandb.ai/deepanshu_rohilla-mrfood-ai/lerobot-piper-clean-v1
Candidate comparison
| Ablation | Step | Train std. MAE | Eval std. MAE | Test std. MAE | Train-test gap | Selected |
|---|
| act-front-top | 15000 | 0.101565 | 0.175120 | 0.191440 | 0.089874 | no |
| act-front-top-left | 10000 | 0.117780 | 0.173132 | 0.191308 | 0.073528 | yes |
| smolvla-front-top | 20000 | 0.130488 | 0.206360 | 0.231821 | 0.101333 | no |
Checkpoint selection within each run used minimum eval standardized MAE. The
cross-run comparison Pareto-minimized eval standardized MAE and absolute
train-test divergence, with their sum as a deterministic frontier tie-break.
Because that criterion inspects the test split, the test result influenced
model-family selection and is not an untouched final audit. A new task-level
dataset or physical rollout campaign is required for a final unbiased audit.
Dataset and preprocessing
- Source:
PranayTest/classified-data-2026-07-06
- Pinned revision:
8ad2dc39b271c3573b2721526f95c10a208c72ba
- Quality JSONL SHA-256:
ff6323b522fb4612c259a5e839c6919619c928957ca929d3bab7110ec5cb343d
- Selected source episodes:
148
- Filter:
class == good, confidence >= 0.9,
camera quarantine [167, 217], and right-arm
command range <= 2.0 degrees
- Cameras:
front, top, left
- Output state/action order:
left_joint1, left_joint2, left_joint3, left_joint4, left_joint5, left_joint6, left_gripper
- Joints: converted from degrees to radians
- Gripper: normalized
[0, 1]
- Actions: absolute commanded targets
- Camera timestamps: excluded from policy inputs
- Split strategy:
chronological_original_episode_order_80_10_10
- The derived 7D-radian datasets were materialized locally from the pinned
source; the exact selection manifest and builder/evaluation code are bundled
with this model release.
| Split | Episodes | Frames |
|---|
| train | 118 | 22424 |
| eval | 15 | 2850 |
| test | 15 | 2849 |
Intended use and limitations
This checkpoint is for offline research and staged robot-policy integration.
The reported errors measure action prediction on recorded demonstrations; they
are not physical task-success rates and do not establish collision safety.
Before any robot execution, use a dedicated inference adapter with fresh-state
checks, controller-level joint/rate/acceleration limits, reject-and-hold
behavior, fault and temperature checks, collision safeguards, an operator
E-stop, and a staged ghost/dry-run/live rollout process. The model is specific
to a left-arm plate pick-and-lift setup and should not be assumed to generalize
to other objects, layouts, cameras, robots, or bimanual behavior.
