3D Brain + Guidance Adapters — submission checkpoints
Checkpoints for the submitted line (Prismatic-0.5B brain + VGGT + 12-site
all_layers_all_blocks guidance adapters, 85.5% mean on 6 tasks x 100
episodes) plus every row of the submission's ablation table, one additional
variant, and two follow-up lines kept working but not part of the submission.
See the project repository's README.md (Results section) and reproduce/
for the full method, per-row provenance, and how to use these.
Each brain_* directory has been trimmed to its single final-epoch
checkpoint (training keeps several for resuming; only the last one is needed
for inference/evaluation). Each dp_* directory has one checkpoint per
(task, arm) expert, already final (dp_ema weights).
Contents
| Directory | Role |
|---|---|
brain_prismatic_B_vggt_v2/ |
Main method brain (multi-depth align, blocks 8/13/18/23) |
dp_gadapt_all12_C150_B50_vggt_v2_6task/ |
Main method DP, 15 of 16 arms |
dp_gadapt_all12_total200_v2val_B50_vggt_v2_3r/ |
Main method DP, ThreeRobotsStackCube arms (headline-table version) |
brain_prismatic_B_vggt_v2_cur/, dp_gadapt_all12_C150_B50_vggt_v2_cur_6task/ |
Ablation: w/o future-alignment (H=0) |
onestage_all12_noalign_vggt_v2_ep150/, onestage_all12_noalign_vggt_v2_ep075_eval/ |
Ablation: w/o alignment (one-stage) |
brain_prismatic_B_vggt_v2_local/, dp_gadapt_all12_C150_B50_vggt_v2_local_6task/ |
Ablation: w/o multi-view input (local brain) |
brain_prismatic_B_da3_v2/, dp_gadapt_all12_C150_B50_da3_v2_6task/ |
Ablation: alignment w/ DA3 |
brain_prismatic_C_vggt_v2/, dp_gadapt_all12_L23_C150_B50_vggt_v2_6task/ |
Additional: single-depth align (block 23 only), not in the submission table |
brain_prismatic_B_dinov2_future_N8_eval6_v2_ep30/, dp_gadapt_all12_C150_B50_dinov2_future_N8_6task/ |
Follow-up, NOT in the submission: DINOv2 as the 3D teacher instead of VGGT/DA3 |
groot_n17_B_vggt_v2_takephoto_all32_extend2k_lr1e5/ |
Follow-up, NOT in the submission: GR00T-N1.7 action expert for 2 of TakePhoto's 4 arms |
sapien_render_580.65.06.sif |
Apptainer/Singularity image for running eval on a headless (no native Vulkan) compute node — see reproduce/README.md's Evaluation section |
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support