Downloads · 30 days
0
MichaelGan0625/piston-n17-ab-15k-diagnostic-20260826
piston-n17-ab-15k-diagnostic-20260826 is a robotics model from MichaelGan0625. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for transformers.
This private diagnostic bundle keeps the two serving checkpoints, their exact LeRobot datasets, teacher-forced open-loop diagnostics, closed-loop simulator rollouts, and the exact Bridge code used for evaluation in on…
Downloads · 30 days
0
Access
Public
Updated Aug 26, 2026
Repo size
15.4 GB
Likes
0
Public
Click a slice to open those files.
.safetensors13.8 GB · 89%
From the Hugging Face model README
This private diagnostic bundle keeps the two serving checkpoints, their exact LeRobot datasets, teacher-forced open-loop diagnostics, closed-loop simulator rollouts, and the exact Bridge code used for evaluation in one place.
The bundle is intended for failure analysis, not as evidence of task success. In particular, the open-loop MP4s replay recorded dataset observations and are not simulator rollouts.
checkpoints/job_A_checkpoint-15000/: Job A serving checkpoint.checkpoints/job_B_checkpoint-15000/: Job B serving checkpoint.datasets/job_A_success87/: 87-episode, 90,275-frame Job A dataset.datasets/job_B_success54/: 54-episode, 43,266-frame Job B dataset.eval/open_loop/: A/B teacher-forced reports and three-camera diagnostic MP4s.eval/closed_loop/: all simulator rollouts, traces, contracts, summaries and MP4s.bridge_eval_snapshot/: exact Bridge/eval code used by the closed-loop runs.TRAINING_RECIPE.md: frozen training contract.EVAL_REPORT.md: evaluation matrix, metrics and current diagnosis.BRIDGE_BRANCH_README.md: code provenance and proposed GitHub branch contents.MANIFEST.json and SHA256SUMS: file inventory and integrity hashes.Job B is not simply underfit. Its teacher-forced error is lower than Job A on the sampled in-distribution anchors, but after a failed grasp it repeatedly issues similar relative right-arm increments. Replanning makes the target ratchet toward joint limits. Higher replanning rates make the failure happen earlier. This is consistent with overfitting to a success-only trajectory and weak state-conditioned recovery, rather than a global camera, transport, or left/right mapping failure.
Review EVAL_REPORT.md and the MP4s before drawing a semantic-success
conclusion.