Downloads · 30 days
8
22% of all-time downloads
prehj/GR00T-N1.5-dexjoco-atq-vlm-labelgated-2.0x-60k-20260919
GR00T-N1.5-dexjoco-atq-vlm-labelgated-2.0x-60k-20260919 is a robotics model from prehj. Use it for the robotics task on the model card, and read the license before you ship it in a product.
Final 60,000-step checkpoint of the DexJoco VLM-label-gated ATQ run, seed 42, global batch 64 (2 GPUs, per-device batch 32). Completed 2026-09-19.
Downloads · 30 days
8
22% of all-time downloads
All-time downloads
36
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
Final 60,000-step checkpoint of the DexJoco VLM-label-gated ATQ run, seed 42, global batch 64 (2 GPUs, per-device batch 32). Completed 2026-09-19.
This is not a DemoSpeedup model. It uses 4 routed experts (main, merged-8, merged-4, native-8), VLM label gating, moe_speed=2.0, conf_threshold=0.5, conf_loss_coef=0.1 and detached confidence readout. DexJoco absolute action targets use action_merge_reduction=last. Labels: prehj/dexjoco-vlm-labels-Astra. Training config: dexjoco_single_arm_conf, embodiment new_embodiment, horizon 16. Requires the matching custom GR00T-action-quantization implementation; a vanilla GR00T loader without these architecture extensions is insufficient.
Model weights, configuration, normalization metadata and trainer progress are included. Optimizer/scheduler/RNG are omitted. This is an inference checkpoint rather than a complete optimizer-resume snapshot. No evaluation success rate is claimed here. File SHA256 values and source/job provenance are provided in publication.json.