Downloads · 30 days
11
32% of all-time downloads
RooibosT/gr00t-n1.7-g1-dex1-ikea-relarm-30hz-h40-v2-bctinit
gr00t-n1.7-g1-dex1-ikea-relarm-30hz-h40-v2-bctinit is a robotics model from RooibosT. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
GR00T N1.7 fine-tuned on carroll511/IKEAtableassembly (v2 conversion) — a Unitree G1 with Dex1 grippers assembling a children's table from a fixed standing position.
Downloads · 30 days
11
32% of all-time downloads
All-time downloads
34
Public
Parameters
3.1B
12.6 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors12.6 GB · 100%
From the Hugging Face model README
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_bctinit / checkpoint-20000.
Trades arm precision for gripper accuracy: gripper -9.6%, arm +1.2%, EE8 +2.2% against the baseline, same sign in all three tasks. Pick this one only if grasp retention fails before arm precision does.
| init | RooibosT/gr00t-n1.7-g1-dex1-bct-relarm-aug-30hz-h40 (round-2 ckpt-15000) |
| modality config | examples/unitree_g1_dex1_ikea/g1_dex1_ikea_relarm_3view_aug_config.py |
| state | 46 — identical key order to the BCT checkpoint, so the state encoder lines up |
| 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) |
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) | 1.455 | 1.457 |
| arm MAE, all 40 steps (deg) | 3.340 | 3.338 |
| wrist position error, steps 1-8 (mm) | 12.10 | 12.11 |
| wrist position error, all 40 (mm) | 20.96 | 20.98 |
| gripper MAE | 0.2083 | 0.2104 |
Per task:
| task | n | arm (deg) | EE8 (mm) | gripper |
|---|---|---|---|---|
| insert table leg to table base | 182 | 2.837 | 13.58 | 0.1568 |
| pick table leg | 180 | 2.862 | 9.36 | 0.1975 |
| rotate leg to tighten | 335 | 3.870 | 12.77 | 0.2420 |
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-h40model's, which was scored on a different split.
get_action() returns unnormalized absolute values.