joyful-stable UR5 GR00T inference checkpoints

This repository contains inference-only GR00T checkpoint exports. Training-only files such as optimizer.pt, scheduler.pt, rng_state.pth, trainer_state.json, training_args.bin, and wandb files are intentionally omitted.

Variants

• v2_filtered/: UR5 hanging 52ea filtered dataset, action[t] = state[t+1].
  • source checkpoint: /home/user/joyful/vla/isaac_gr00t/finetuned_models/ur5_hanging_v2_0530_310_430mm_52ea_filtered_comp_hanging/checkpoint-15000
  • source dataset: /home/user/joyful/vla/datasets/lerobot_2.1.0/ur5_hanging_v2/0530_310-430mm_52ea_filtered/comp_hanging
• v3_gripper_dense/: UR5 hanging 52ea gripper-dense variant. Gripper scaled to 0.0..0.7 and gripper action[t] = processed state[t+2].
  • source checkpoint: /home/user/joyful/vla/isaac_gr00t/finetuned_models/ur5_hanging_v3_gripper_dense_0530_310_430mm_52ea
  • source dataset: /home/user/joyful/vla/datasets/lerobot_2.1.0/ur5_hanging_v2/0530_310-430mm_52ea_gripper_dense_v3/comp_hanging
• v4_gripper_dense_7p5k/: UR5 hanging 52ea gripper-dense variant. Same data/config as v3, trained for 7.5k steps.
  • source checkpoint: /home/user/joyful/vla/isaac_gr00t/finetuned_models/ur5_hanging_v4_gripper_dense_0530_310_430mm_52ea_7p5k
  • source dataset: /home/user/joyful/vla/datasets/lerobot_2.1.0/ur5_hanging_v2/0530_310-430mm_52ea_gripper_dense_v3/comp_hanging
• v5_gripper_dense_0_1_7p5k/: UR5 hanging 52ea gripper-dense variant. Same filtered source as v2, gripper scaled per episode to 0.0..1.0, trained for 7.5k steps.
  • source checkpoint: /home/user/joyful/vla/isaac_gr00t/finetuned_models/ur5_hanging_v5_gripper_dense_0_1_0530_310_430mm_52ea_7p5k
  • source dataset: /home/user/joyful/vla/datasets/lerobot_2.1.0/ur5_hanging_v2/0530_310-430mm_52ea_gripper_dense_v5/comp_hanging
• v6_gripper_dense_15k/: UR5 hanging gripper-dense v6 (42ep, end-plateau filtered from v5). Gripper scaled per episode to 0.0..1.0, gripper action[t] = processed state[t+2], arm action[t] = state[t+1]. Trained for 15k steps.
  • source checkpoint: /home/user/joyful/vla/isaac_gr00t/finetuned_models/ur5_hanging_v6_gripper_dense_0530_310_430mm_52ea_15k
  • source dataset: /home/user/joyful/vla/datasets/lerobot_2.1.0/ur5_hanging_v2/0530_310-430mm_52ea_gripper_dense_v6/comp_hanging
• v6_gripper_dense_30k/: UR5 hanging gripper-dense v6 (42ep, end-plateau filtered from v5). Gripper scaled per episode to 0.0..1.0, gripper action[t] = processed state[t+2], arm action[t] = state[t+1]. Trained for 30k steps.
  • source checkpoint: /home/user/joyful/vla/isaac_gr00t/finetuned_models/ur5_hanging_v6_gripper_dense_0530_310_430mm_52ea_30k
  • source dataset: /home/user/joyful/vla/datasets/lerobot_2.1.0/ur5_hanging_v2/0530_310-430mm_52ea_gripper_dense_v6/comp_hanging
• v7_1_smooth_30ep/: UR5 hanging 0606 v3 (106ep, merged from 3 raw sets, anomalous + blue-cast episodes dropped). Arm return-to-home step smoothed to action[t]=state[t+4]; gripper kept RAW. Trained ~30 epochs.
  • source checkpoint: /home/user/joyful/vla/isaac_gr00t/finetuned_models/ur5_hanging_v3_smooth_0606_111ea_30ep
  • source dataset: /home/user/joyful/vla/datasets/lerobot_2.1.0/ur5_hanging_v3/fixed/comp_hanging_v3
• v7_2_gripfix_15ep/: UR5 hanging 0606 v3 (106ep), same arm smoothing as v7, plus gripper per-episode affine rescale [ep_min,ep_max]->global [min,max] (state & action separately) to remove per-episode open/close spread while preserving trapezoid shape and close timing. Trained ~15 epochs.
  • source checkpoint: /home/user/joyful/vla/isaac_gr00t/finetuned_models/ur5_hanging_v8_gripfix_0606_111ea_15ep
  • source dataset: /home/user/joyful/vla/datasets/lerobot_2.1.0/ur5_hanging_v3/fixed/comp_hanging_v3_gripfix
• v8_1_n17_with_v_6_30k/: Identical data/config/preprocessing to v6_gripper_dense_30k (UR5 hanging 42ep, gripper scaled per episode to 0.0..1.0, gripper action[t] = processed state[t+2], arm action[t] = state[t+1], 30k steps), but trained with GR00T 1.7 instead of 1.6.
  • source checkpoint: /home/user/joyful/vla/isaac_gr00t_n17/finetuned_models/ur5_hanging_v6_gripper_dense_0530_310_430mm_52ea_30k_n17
  • source dataset: /home/user/joyful/vla/datasets/lerobot_2.1.0/ur5_hanging_v2/0530_310-430mm_52ea_gripper_dense_v6/comp_hanging

Download for local inference

Example for the gripper-dense variant:

huggingface-cli download pilab-rfm/joyful-stable --include 'v3_gripper_dense/*' --local-dir ./joyful-stable
MODEL_PATH=$PWD/joyful-stable/v3_gripper_dense

Then pass MODEL_PATH to the existing prism/vla_inference/server/run_server.sh flow.