UMC teachers and RGB decoders
Private inference weights. Optimizer states are not included. Absolute local paths were stripped from configs.
UMC teachers
| file | domain | step | contract |
|---|---|---|---|
umc_stage_a_aligned4_step080000.inference.pt |
LIBERO-X | 80000 | 8/8 fusion-decoder, pixel-flow, RAE joint, DA3 aligned4, 2 views, H64 stratified |
umc_robocasa_aligned4_step030000.inference.pt |
RoboCasa Human300 | 30000 | same 80k architecture, 3 views, Sea-RAFT, image-only DA3 aligned4, warm-start from the 80k teacher |
umc_robocasa_aligned4_step080000.inference.pt |
RoboCasa Human300 | 80000 | continuation of the 30k Human300 teacher to 80k; 3 views, H4–H32 |
umc_simpler_bridge_aligned4_step030000.inference.pt |
SimplerEnv Bridge / WidowX | 30000 | same 80k architecture, 1 view (image_0 / 3rd_view_camera), H1–H8 at 5 FPS |
Load with torch.load(..., weights_only=False)['model']. Formats: umc_stage_a.training.v2, umc_robocasa.training.v2, umc_simpler.training.v1.
Normalization JSONs:
normalization_h64_aligned4_all_unclipped.json— aligned4 / RoboCasa / Bridgenormalization_8x8_pixel_flow.json— earlier pixel-flow Stage-A
RGB decoders (visualization only)
RAE patch-pixel decoders mapping frozen DINO / V-JEPA tokens to RGB. Not UMC teachers.
| file | teacher | domain | notes |
|---|---|---|---|
rgb_decoder_dino_libero_30k.pt |
DINO | LIBERO | 30k pixel-L1 |
rgb_decoder_vjepa21_libero_30k.pt |
V-JEPA 2.1 | LIBERO | 30k pixel-L1 |
rgb_decoder_dino_human300_from30k.pt |
DINO | Human300 | continued from the LIBERO frozen decoder |
rgb_decoder_vjepa21_human300_from30k.pt |
V-JEPA 2.1 | Human300 | continued from the LIBERO frozen decoder |
rgb_decoder_dino_bridge_from_human300.pt |
DINO | Bridge / WidowX image_0 |
30k pixel-L1 from the Human300 frozen decoder |
rgb_decoder_vjepa21_bridge_from_human300.pt |
V-JEPA 2.1 | Bridge / WidowX image_0 |
30k pixel-L1 from the Human300 frozen decoder |
Format: umc_stage_a.rgb_decoder.frozen.v1. Load with umc_stage_a.rgb_decoder.load_frozen_rgb_decoder.