Downloads · 30 days
11
32% of all-time downloads
easyminnn/atq-labelgated-2.0x-openarm-vlmconf-v2-40k
atq-labelgated-2.0x-openarm-vlmconf-v2-40k is a robotics model from easyminnn. 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 other.
Inference-only checkpoint (weights + config; optimizer/scheduler/RNG state stripped).
Downloads · 30 days
11
32% of all-time downloads
All-time downloads
34
Public
Parameters
2.8B
8 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors8 GB · 100%
How the weights are stored.
BF161.7B · 58%
From the Hugging Face model README
Inference-only checkpoint (weights + config; optimizer/scheduler/RNG state stripped).
Fine-tuned from nvidia/GR00T-N1.5-3B on the openarm bimanual teleop set
(107 episodes, 66,203 frames, 30 fps, 2 egocentric cameras, 26-dim absolute joint
targets: left_arm 0:7, right_arm 7:14, left_hand 14:20, right_hand 20:26) with the
VLM speed-ratio labels of prehj/openarm-conf-labels-v2 baked into the action tensor
(ratio_label, dims 26:28; stride-16 anchors linearly interpolated to every frame).
| Experts (4) | main h16 1x · n8 h8 1x · m8 h16 2.0x (8 rows) · m4 h8 2.0x (4 rows) |
| Merge reduction | last (absolute joint targets: a merged block takes its last value) |
| Gate | conf >= tau opens the compressed group; router ranks horizons inside it. tau is an eval knob (checkpoint value 0.65) |
| Data config | openarm_bimanual_conf (GR00T-action-quantization, branch jimin-dev-label-gated) |
| Training | global batch 64 (32 x 2 H100), 40k steps, lr 1e-4, conf_loss_coef 0.1, conf readout detached, router warmup 5000 |
| Embodiment tag | new_embodiment |
Training diagnostics at the end of the run: conf MAE ~0.03 against the label; router probabilities stayed near-uniform across the four experts (comp_split ~0.5), so horizon choice inside a group is close to random — the conf gate is the active part.
Serve with scripts/inference_service.py --head moe --data-config openarm_bimanual_conf --embodiment-tag new_embodiment [--conf-threshold TAU].