Downloads · 30 days
5
13% of all-time downloads
sreetz-nv/so101_lior70_h32_lr1e4_20k_20260915
so101_lior70_h32_lr1e4_20k_20260915 is a robotics model from sreetz-nv. 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 lerobot.
Final 20,000-step checkpoint trained on the 70-episode Lior teal dataset. The policy uses 32-action chunks and 32 executed action steps per query.
Downloads · 30 days
5
13% of all-time downloads
All-time downloads
40
Public
Parameters
3.1B
12.6 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors12.6 GB · 100%
From the Hugging Face model README
Final 20,000-step checkpoint trained on the 70-episode Lior teal dataset. The policy uses 32-action chunks and 32 executed action steps per query.
nvidia/GR00T-N1.7-3B.liorbenhorin-nv/so101_vials_sysid_dr_teal_70ep.b4bb7a8576255038a7a9f09d1ae7551958c05df9.Full settings are in train_config.json. Its filesystem paths describe the
training machine and need to be overridden for another environment.
RGB inputs are observation.images.external and observation.images.wrist.
State/actions use the workshop joint order: shoulder pan, shoulder lift, elbow
flex, wrist flex, wrist roll, gripper. Language instruction:
"Pick up the vial and place it in the rack".
This model uses the dataset's legacy workshop range-scaled joint convention.
On a workshop version with explicit mapping selection, evaluate with
--joint_mapping workshop. It was not trained on the new degree-preserving
recordings. Direct real-robot degree passthrough does not preserve this convention;
physical alignment and a consistent state/action conversion require validation.
Use the saved LeRobot pre/postprocessors for image/state processing and absolute
action reconstruction.
On September 15, 2026, this checkpoint completed 13/20 episodes in the workshop vials-to-rack evaluation: teal robot, raised starting pose, fresh camera images, 30-second simulated episode limit, H32, no added target hold, environment seed 42. All 402 inference queries returned 32 actions. The supplied H16 reference scored 12/20 under the same nominal protocol; this small, unpaired comparison does not establish a longer-horizon advantage. This is a simulation result, not a validated real-robot success rate.