pashuparthis/mimic_apple_pick_and_place

G1 + BrainCo Revo2 apple pick-and-place — GR00T-LeRobot fine-tuning data

Isaac Lab Mimic demonstrations of a Unitree G1 picking an apple and placing it on a plate, recorded on the BrainCo Revo2 Touch hands rather than the Dex3-1 the source policy was trained on. These are the datasets used to fine-tune GR00T N1.7 back across that embodiment gap, in GR00T-LeRobot v2.1 format — ready to hand to gr00t/experiment/launch_finetune.py as-is.

Config	Episodes	Frames	Size	What it is
r1r2	1,668	821,610	1.8 GB	Rounds 1–2. Apple spawn jittered uniformly in a ±5 cm box.
r3	979	483,294	1.0 GB	Round 3. Targeted at the spawn cells the r1r2 policy failed in.
r2r3	2,647	1,304,904	2.7 GB	The union, and the one the final checkpoint was trained on.

r2r3 is exactly r1r2 + r3 (2647 = 1668 + 979 episodes, 1,304,904 = 821,610 + 483,294 frames), so download one, not all three unless you want to reproduce the ablation between them.

Why three datasets and not one

The source policy (nvidia/GN1x-Tuned-Arena-G1-Static-PickNPlace) scores 0.65 on the Dex3-1 hands it was trained on and 0.06 on the Revo2 hands, after an analytic retargeting layer had matched the two hands about as well as a closed-form map can. The residual gap was behavioural — a five-finger cage around an opposed thumb holds an apple differently from a three-finger pinch, and the Revo2 takes 0.65 s to close against the Dex3's 0.2 s — so the fix had to be demonstrations.

None were teleoperated. The 6% of episodes that already succeeded were harvested from the retargeted policy itself, and Isaac Lab Mimic replayed them against randomised scenes:

• Rounds 1–2 (r1r2) scaled 22 source demos into 1,668 by jittering the apple spawn uniformly. Fine-tuning on this took the task from 0.06 to 0.95 at the original deterministic spawn.
• That policy then scored 0.50 at ±5 cm and 0.00 at ±10 cm — it had learned the task but not the workspace. Round 3 (r3) gridded the spawn box, measured success per cell, and generated only where the policy was failing, instead of sampling uniformly again.
• r2r3 merges them. Fine-tuning on it reached 0.55 at ±5 cm (n=40) — statistically indistinguishable from the 0.50 that r1r2 alone achieved (p = 1.0), which is a result worth publishing alongside the data rather than hiding: targeted generation did not buy what it was supposed to buy at this scale.

Schema

GR00T-LeRobot v2.1: data/ (parquet), videos/ (MP4), meta/ (info.json, modality.json, episodes.jsonl, tasks.jsonl, stats.json).

Key	Type	Shape
observation.images.ego_view	video, 50 fps	480 × 640 × 3
observation.state	float32	43
action	float32	43
action.eef_pose, observation.eef_pose	float32	14 (both wrists, pos + quat)

Single task string: "move the apple to the plate". One camera, the robot's head camera — there is no wrist camera and no proprioceptive history, so the policy is Markovian on a single frame.

meta/modality.json splits the 43 dims as left_leg 0–6, right_leg 6–12, waist 12–15, left_arm 15–22, left_hand 22–29, right_arm 29–36, right_hand 36–43. The task is single-handed: only the left arm and hand move, and left_hand dim 23 is a clean binary grasp signal (0.0 open, −0.87 rad closed) that is useful for auditing.

One thing to know before training on this

Every one of the 2,647 episodes closes the hand exactly once. There are no failed-then-retried grasps anywhere in the data, because the episodes were success-filtered by construction. A policy behaviour-cloned on this cannot recover from a missed grasp, and no inference-time trick recovers it either — action-chunk ensembling and real-time chunking were both tried and neither produced a retry. If you need recovery behaviour, it has to come from data that contains it.

Use

from datasets import load_dataset
ds = load_dataset("pashuparthis/mimic_apple_pick_and_place", "r2r3", split="train")

Fine-tuning GR00T N1.7 directly on the tree:

hf download pashuparthis/mimic_apple_pick_and_place --repo-type dataset \
  --include 'r2r3/*' --local-dir ./apple_data

python -m torch.distributed.run --nproc_per_node=3 --standalone \
  gr00t/experiment/launch_finetune.py \
    --base-model-path nvidia/GN1x-Tuned-Arena-G1-Static-PickNPlace \
    --dataset-path ./apple_data/r2r3 \
    --modality-config-path .../g1_sim_wbc_data_gr00t_n_1_7_config.py \
    --embodiment-tag new_embodiment \
    --global-batch-size 240 --max-steps 10000 --learning-rate 1e-4 \
    --no-tune-llm --tune-visual --tune-projector --tune-diffusion-model

Generation

Scene	galileo_g1_static_apple_mimic (static-base G1)
Recording embodiment	g1_wbc_agile_pink_brainco (Pink IK, Revo2 hands)
Deployment embodiment	g1_wbc_agile_joint_brainco (43-D joint targets, Revo2 hands)
Object → destination	apple_01_objaverse_robolab → clay_plates_hot3d_robolab
Randomisation	per-episode apple spawn jitter, per-worker seed, per-episode grasp depth
Control rate	50 Hz

Mimic is run on a Pink-IK embodiment wearing the Revo2 hands, which expands the hand command through the same retargeting layer used at deployment. The grasp recorded in the data is therefore the grasp the deployed policy actually makes, not a Dex3 grasp relabelled.

The raw per-worker HDF5 shards that earlier versions of this repo hosted have been removed; these LeRobot trees are their converted, merged and validated form, which is what the training code consumes.

Provenance and licences

Derived from NVIDIA IsaacLab-Arena release/0.2.1 (Apache 2.0) and the nvidia/GN1x-Tuned-Arena-G1-Static-PickNPlace checkpoint (NVIDIA Open Model Licence), on Unitree G1 and BrainCo Revo2 assets supplied under their own terms. The demonstrations themselves are released Apache 2.0. Entirely simulation — no hardware, no human teleoperation.