Ronaldo-GOAT/jh_data_aug

Pose6DAug · sim action-augmented dataset + GR00T-N1.5 checkpoints

One self-contained repo: the 1920-episode action-augmented dataset, the checkpoints (base + resume), the training configs, and a guide to reproduce or resume the run.

• Task: "pick up the object and place it in the white basket" (7-DoF arm + gripper)
• Base model: NVIDIA GR00T-N1.5-3B
• Embodiment tag: new_embodiment · Data config: real_droid_joint (absolute-joint recipe)
• Training code: Ronaldo-GOAT/train_mygr00t

huggingface-cli download Ronaldo-GOAT/jh_data_aug --repo-type dataset --local-dir jh_data_aug

---

What "sim action-augmentation" is

Each augmented episode starts from a real teleoperation recording (7-DoF arm + gripper). The recorded arm motion (joints 1–6) is replayed in a MuJoCo scene so the composited video shows the real arm; only around the grasp does a policy take over joint 7 (wrist) + the gripper, in a symmetric window of 10 frames before → 10 frames after the recorded grasp frame ("buf10", a ~1 s buffer each side). Before that window the recorded actions are reused; after it the arm is replayed back onto the recorded schedule.

On top of that, each config applies object-placement jitter: 8 base orientations, yaw jitter (±5° "b" configs / ±10° "w" configs), random-direction translation up to 3 cm, and object scale jitter (0.96–1.04, baked into the scene mesh). Objects are re-targeted across a pool (e.g. a recorded blue-cup grasp drives an abc_chocolate / mango / banana / red_cleanser target where grasp height is feasible).

Result: 1920 episodes over 4 target objects (480 each): abc_chocolate, banana, mango, red_cleanser.

---

Dataset

Format	LeRobot v2.0 (codebase_version: v2.0)
Episodes	1920
Frames	396,026
Videos	3840 (2 views × 1920 episodes)
FPS	10 · Robot type franka

All episodes live in data/chunk-000/ and videos/chunk-000/; meta/info.json sets chunks_size to total_episodes + 1 (1921) so that episode_chunk = episode_index // chunks_size is 0 for every episode_index and the whole dataset resolves to chunk-000.

Observation / action space (real_droid_joint)

• Videos: observation.images.exterior_image_1_left, observation.images.wrist_image_left — 448×252 h264, 10 fps, one frame per row.
• State (8): joint_pos_abs[7] + gripper_close[1].
• Action (8): joint_pos_abs[7] + gripper_close[1], 16-step action chunk.
• Joints 1–6 are the recorded arm; joint 7 + gripper come from the rollout inside the augmentation window. Normalization: min-max (stats in configs/norm_stats_metadata.json).

The 960-episode training subset

The released policy was trained on a 960-episode subset (240 random episodes per target object, seed 42) — not a separate dataset. Every subset episode is a byte-identical copy of an episode here, with only episode_index / index renumbered to 0..959. It is defined by subset_960_indices.txt (960 episode_index values) and described in SUBSET_960.md.

Rebuild it exactly with the shipped script:

python jh_data_aug/resume/scripts/make_subset_960.py \
  --parent jh_data_aug \
  --meta   jh_data_aug/resume/subset_960_meta \
  --out    subset_960 --link

The subset carries its own meta/ — notably meta/stats.json, the min-max stats computed over 960 episodes, which differ from the 1920-episode stats. That meta ships in resume/subset_960_meta/, and the script drops it in, so the result is bit-identical to the directory the run was trained on.

---

Contents

meta/ data/ videos/               the 1920-episode LeRobot dataset
subset_960_indices.txt            episode_index values of the 960-episode training subset
subset_960.json                   same, plus name / source / target / cfg / length per episode
SUBSET_960.md                     how the subset is defined and rebuilt
episode_source_map.json           episode_index -> source episode / target / cfg (1920)
conversion_stats.json             builder statistics

checkpoint-30000-base/            base policy: GR00T-N1.5 fine-tuned on 120 real episodes, 30k steps
checkpoint-10000-jhaug960-resume/ action-augmented run at step 10k/80k, with FULL trainer state
                                  (optimizer + scheduler + RNG) for an exact resume

configs/
  train_960.sh                    original training launcher
  sel_1920.txt                    the 1920 (epkey cfg) selections
  sel_960.txt                     the 960-episode training subset selection (240/object, seed 42)
  code/p24_build_lerobot_buf10.py dataset builder (frames -> LeRobot, abs joints)
  norm_stats_metadata.json        min-max normalization stats (new_embodiment)

resume/
  RESUME.md                       step-by-step "continue this run elsewhere" guide
  env/requirements-mygr00t.txt    pip freeze of the training env
  env/conda-mygr00t.txt           conda list --export of the same env
  env/versions.txt                python / torch / CUDA / GPU summary
  scripts/make_subset_960.py      rebuild the 960-episode training subset
  scripts/train_960_2gpu.sh       2 GPUs x batch 32  (the config that produced step 10k)
  scripts/train_960_4gpu.sh       4 GPUs x batch 16
  scripts/train_960_4gpu_numa.sh  4 GPUs with per-rank numactl binding
  scripts/rank_wrap.sh            per-rank NUMA wrapper used by the script above
  subset_960_meta/                the subset's own meta/ + episode_source_map.json

---

Reproduce (fresh 80k run from the base checkpoint)

Global batch 64 = 32 × 2 GPUs.

huggingface-cli download Ronaldo-GOAT/jh_data_aug --repo-type dataset --local-dir jh_data_aug
python jh_data_aug/resume/scripts/make_subset_960.py \
  --parent jh_data_aug --meta jh_data_aug/resume/subset_960_meta --out subset_960 --link

torchrun --standalone --nnodes=1 --nproc_per_node=2 gr00t_finetune.py \
  --dataset-path subset_960 \
  --output-dir runs/gr00t_n15_jhaug960_newemb_droid_joint_gbs64_80k_from30k \
  --dataloader-num-workers 16 --data-config real_droid_joint \
  --embodiment-tag new_embodiment \
  --base-model-path jh_data_aug/checkpoint-30000-base \
  --run-name jhaug960_gbs64_80k_from30k \
  --video-backend torchvision_av \
  --batch-size 32 --num-gpus 2 \
  --save-steps 5000 --save-at 5000,10000,20000,30000,40000,50000,60000,70000,80000 \
  --prediction-mode base --stage1-step 80000

Training code: the GR00T finetuning stack at Ronaldo-GOAT/train_mygr00t (PYTHONPATH=$TN/MoSS_GR00T, IS_TORCHRUN=1).

Variants: the released policy trains on the 960 subset (240 / object); the full 1920 episodes are here for training larger runs.

---

Resume on another server

The in-progress 80k run (jhaug960_gbs64_80k_from30k) can be continued on a different machine from step 10,000 / 80,000, with the optimizer, LR scheduler and RNG state intact — an exact continuation, not a warm restart.

1. huggingface-cli download Ronaldo-GOAT/jh_data_aug --repo-type dataset --local-dir <DS>, then build the 960-episode training subset with resume/scripts/make_subset_960.py.
2. Download checkpoint-10000-jhaug960-resume/ from this repo and place it as <output-dir>/checkpoint-10000 — the trainer resumes from the last checkpoint found in --output-dir when --stage1-resume is passed, and the step is read from the folder name, so the name must be exactly checkpoint-10000.
3. Get the code from Ronaldo-GOAT/train_mygr00t and rebuild the env from resume/env/.
4. Run the same torchrun command as above plus --stage1-resume, keeping global batch = 64 (2 GPUs × 32, or 4 GPUs × 16).

--base-model-path still points at checkpoint-30000-base/ (it supplies the config); the weights/optimizer come from checkpoint-10000. Resuming with a different per-device batch prints a benign per_device_train_batch_size mismatch warning — the global batch is unchanged. ~70,000 steps remain; saves land at 20k/30k/40k/50k/60k/70k/80k.

Full details, exact commands and caveats: resume/RESUME.md.