Modelpublic

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

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes8downloads

Model Card

IKEA v2 — baseline, 30k schedule

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_30k / checkpoint-26000 (best of 15 scanned checkpoints).

Best checkpoint of the 30k cosine schedule. Versus the 20k-schedule model (...-h40-v2): gripper -6.4%, arm/EE unchanged — only the gripper improves past 20k, flattening at 26k. No arm-velocity input.

Setup


init	`nvidia/GR00T-N1.7-3B` (backbone frozen: `tune_llm=False`, `tune_visual=False`)
modality config	`examples/unitree_g1_dex1_ikea/g1_dex1_ikea_relarm_3view_aug_config.py`
state	46 (legs, waist, leftarm, rightarm, leftgripper, rightgripper, basegravity, lefteef, right_eef)
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, 30,000 steps (cosine, warmup 5%), 2x A100 DDP

Open-loop accuracy

Held-out _v2_val split, 697 windows at stride 10, 4 denoising steps, fixed per-window seed.

metric	checkpoint-26000	mean of 26k/28k/30k
arm MAE, steps 1-8 (deg)	1.450	1.451
arm MAE, all 40 steps (deg)	3.260	3.258
wrist position error, steps 1-8 (mm)	11.99	11.96
wrist position error, all 40 (mm)	20.76	20.69
gripper MAE	0.2168	0.2179

Per task:

task	n	arm (deg)	EE8 (mm)	gripper
insert table leg to table base	182	2.705	13.72	0.1482
pick table leg	180	2.851	9.44	0.1927
rotate leg to tighten	335	3.781	12.43	0.2671

Read the 1-8 step figures: async deployment executes ~8 steps after latency compensation, so late chunk steps never run. eval_loss does not track action quality.

These numbers are from the v2 val split — not comparable to ...-h40 (v1) models.

Caveats

—Open-loop only; closed-loop behaviour on hardware unverified for this checkpoint.
—Arms are RELATIVE, grippers ABSOLUTE; get_action() returns unnormalized absolute values.
—The waist is not commanded; recordings are from a fixed stance (waist pitch bimodal at ~2.6/10.4 deg — set the robot near 10.5 deg, the majority mode).