right_pnp_multicube_32b_05d_175eps

symbiosika/right_pnp_multicube_32b_05d_175eps

No description was written for this artifact. Everything below is derived from the files themselves, so it is accurate but says nothing about intent, context or known problems. To add that, write mlops/cards/right_pnp_multicube_20260826-173350.md and publish again.

This card reports training, not performance. No on-hardware success rate is claimed anywhere below. Loss is not a success rate; a policy with a clean loss curve can still fail every rollout. The way this repo answers did it work is bc/eval.py --record followed by python -m dataset_critic <rollout> --rollout, and that verdict lives next to the rollout, not here.

Training

Training data

The dataset was local at training time and is not linked here. Publish it with python -m mlops publish dataset <path> and re-publish this model with --dataset-repo <id> to link the two.

Loss

Last metrics line as lerobot logged it:

Training: 100%|██████████| 20000/20000 [15:29:10<00:00,  2.77s/step]INFO 2026-08-27 09:03:05 ot_train.py:641 step:20K smpl:640K ep:2K epch:9.25 loss:0.010 grdn:0.104 lr:2.1e-09 updt_s:2.521 data_s:0.247 smp/s:12 mem_gb:36.23

No validation loss: this run held no episodes out (--eval-split 0), so the loss above is training loss only and says nothing about generalization.

Weights

These are lerobot's float32 masters. For a model whose weights are bf16 throughout, that is roughly twice its resident size, and _load_as_safetensor puts the whole state dict on the GPU before copying it into the parameters -- so loading this file needs VRAM equal to its size. bc/to_bf16.py halves it.

Evaluating this checkpoint

huggingface-cli download symbiosika/right_pnp_multicube_32b_05d_175eps --local-dir ./right_pnp_multicube_20260826-173350

# load and verify against the live rig WITHOUT energizing anything:
python bc/eval.py --checkpoint ./right_pnp_multicube_20260826-173350 --check

Two prerequisites that are easy to miss:

• Calibration is part of the checkpoint. A joint-space policy acts in the frame defined by the follower arm's calibration file. Recalibrating between training and evaluation shifts every state and action silently, and the policy starts grasping centimeters off target while looking healthy otherwise.
• The base is a local path, baked into this checkpoint. Its config names base_models/GR00T-N1.7-3B-custom/, and that directory is loaded before these weights are. On another machine the same path has to exist, or the config has to be edited to point at wherever the base actually is — the Hub id in this card's metadata is the lineage, not a drop-in substitute.

Files

• config.json
• model.safetensors
• policy_postprocessor.json
• policy_postprocessor_step_0_groot_action_unpack_unnormalize_v2.safetensors
• policy_preprocessor.json
• policy_preprocessor_step_2_groot_n1_7_pack_inputs_v1.safetensors
• train.log
• train_config.json

---

Card generated by mlops/publish.py from the checkpoint's own config, its train_config.json and its train.log; the description above it is hand-written.