sreetz-nv/so101_vials_procedural250_sim75_gr00t17_h16_20k_20261011

SO-101 procedural and simulated-teleoperation policy

This is a fresh GR00T N1.7 policy trained to test whether combining varied scripted demonstrations with simulated teleoperation improves vial pickup and rack-placement behavior. It was trained by Shane Reetz from 250 procedural simulation episodes and 75 simulated-teleoperation episodes. It contains no real-robot training data.

Training data

• liorbenhorin-nv/so101_vials_dr_procedural_layout_aware_250ep_20261009, revision b9018e04401a8dee9fd777908db777760a8bb64f: 250 scripted episodes with procedural layouts and domain randomization.
• sreetz-nv/so101_sim_manual_perception_newton_20261008, revision 4137a62afa9758f5f6c4b520d9e691e24f492977: 75 simulated-teleoperation episodes.
• Combined training set: 325 episodes and 113,200 frames, sampled at their natural frame frequency.

The merged export was audited for exact numeric equality with the two source datasets and copied-video hash equality.

Training configuration

• Base model: nvidia/GR00T-N1.7-3B, revision 2fc962b973bccdd5d8ce4f67cc63b264d6886495
• Optimizer updates: 20,000
• Supervised action horizon: 16
• Effective global batch: 64, implemented as microbatch 16 with four accumulation steps on one GPU
• Learning rate: 1e-4; 1,000 warmup updates; cosine schedule
• Vision backbone: trainable
• Language backbone: frozen
• Seed: 42
• Training implementation revision: 30da8e687a6dfc617fcd94afc367ac7071c376ce
• Recipe document revision: d7ef9bf25b3880d4c5cd6f54f52711098f5d286f

The 20,000-update receipt recorded a final learning rate of zero. The exported model includes its policy preprocessor and postprocessor.

Evaluation status

Simulation and physical-robot evaluation results are not included yet. Evaluation should use an exact recorded starting pose from the relevant source-data group and record the achieved pose before policy actions. The scripted and teleoperation subsets have different starting-pose distributions; a single start does not establish performance across both groups.

For physical use, verify joint order and units, gripper calibration, camera identity, action execution horizon, runtime revision, and safety limits. These weights do not establish safe or reliable operation on a physical robot.

Limitations

This checkpoint was trained only with simulation observations and actions. It does not by itself demonstrate real-robot transfer, robustness to unseen starts, recovery behavior, or repeated-cycle reliability. Results obtained from a start outside the training-data distributions should be reported separately as out-of-distribution trials.