Modelpublic

RooibosT/gr00t-n1.7-g1-dex1-ikea-relarm-30hz-h40-v2-armvel

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes13downloads

Model Card

IKEA v2 — arm velocity

GR00T N1.7 fine-tuned on carroll511/IKEA_table_assembly (v2 conversion) — a Unitree G1 with Dex1 grippers assembling a children's table from a fixed standing position.

g1_dex1_ikea_relarm_3view_aug_b64_v2_armvel / checkpoint-20000.

Best of the three over the window async deployment actually executes: arm -16.7% and EE -9.2% at steps 1-8, still -7.3% / -6.0% at steps 9-16. Requires arm joint velocities at inference. Feeding zeros instead costs +26% at steps 1-8, which lands 5.4% worse than the baseline that never had velocity.

Setup


init	`nvidia/GR00T-N1.7-3B` (stock)
modality config	`examples/unitree_g1_dex1_ikea/g1_dex1_ikea_armvel_config.py`
state	60 — the 46-dim set plus `left_arm_vel` and `right_arm_vel` (14 dims, inserted after `right_arm`)
action	16 dims — `left_arm`(7, RELATIVE), `right_arm`(7, RELATIVE), `left_gripper`, `right_gripper`
horizon	40 @ 30 Hz
cameras	`cam_left_high`, `cam_left_wrist`, `cam_right_wrist`
training	effective batch 64 (global 16 x accum 4), 20,000 steps, 2x A100 DDP
backbone	frozen (`tune_llm=False`, `tune_visual=False`)

Open-loop accuracy

Scored on the held-out _v2_val split (26 episodes split at session boundaries, 697 windows at stride 10, 4 denoising steps, fixed per-window seed).

metric	checkpoint-20000	mean of 16k/18k/20k
arm MAE, steps 1-8 (deg)	1.206	1.207
arm MAE, all 40 steps (deg)	3.151	3.152
wrist position error, steps 1-8 (mm)	10.74	10.77
wrist position error, all 40 (mm)	20.03	20.08
gripper MAE	0.2182	0.2191

Per task:

task	n	arm (deg)	EE8 (mm)	gripper
insert table leg to table base	182	2.649	12.75	0.1595
pick table leg	180	2.710	8.25	0.1858
rotate leg to tighten	335	3.661	10.98	0.2676

Read the 1-8 step figures: async deployment executes about 8 steps after latency compensation and spends the rest of the chunk on RTC overlap, so steps past ~16 never run. eval_loss does not track action quality here and must not be used to pick a checkpoint.

These numbers come from the v2 val split and are not comparable to the earlier ...-30hz-h40 model's, which was scored on a different split.

Caveats

—Open-loop only. Closed-loop behaviour on hardware is unverified.
—The action space is RELATIVE for the arms and ABSOLUTE for the grippers; get_action() returns unnormalized absolute values.
—The waist is not commanded (the recordings are all from a fixed stance) and there is no locomotion signal in this dataset.