Add Cosmos3-Edge variant of the LIBERO-10 action-policy SFT recipe

[filipemartinsubrobotics]

What

action_policy_libero_edge: the libero_10 action-policy SFT recipe on the public
nvidia/Cosmos3-Edge base. It mirrors action_policy_libero_nano and differs only in

• the model config (EDGE_MODEL_CONFIG, dense Nemotron-2B-Dense-VL backbone), and
• one extra trainable key, k_norm_und_for_gen, as in vision_sft_edge.

Adds the experiment, a run TOML and launcher mirroring Nano preset A (HSDP 2x8,
global batch 2048), and a docs section covering the one-node setting
(replicate 1, grad_accum 2) and running outside the Docker image.

Validation

• 50-iteration single-node smoke run, 8x RTX PRO 6000 Blackwell 96 GB, global batch 2048:
  ~79 s/iteration, loss 15.3 -> 3.6, DCP checkpoint saved.

• Full 2000-iteration run of this recipe (same experiment, lr, schedule and global batch 2048)
  on 4x RTX PRO 6000 Blackwell 96 GB: shard 4, replicate 1, grad_accum 4; ~220-310 s/iteration,
  final loss 0.90.

• Closed-loop libero_10 eval of iter_000002000, following the docs (section 3): server with
  --action-normalization quantile_rot, the bundled libero_native_frame_wise_relative_rot6d.json
  stats, --raw-action-dim 10 --fps 20, 30 sampling steps; client --camera agentview,wrist --image_size 256 --action_space frame_wise_relative --rotation_space 6d --action_dim 10,
  50 trials per task, 8 envs, seed 0.

  Success rate: 407/500 = 81.4 %

Closed-loop LIBERO-10 rollouts (final checkpoint iter_000002000)

libero10_cosmos3_edge_iter2000.mp4

_{One episode per libero_10 task, agent-view camera, instruction shown on top. The 81.4 % above comes from the separate 500-episode run (50 trials per task). Full page: https://huggingface.co/spaces/ubr-physical-ai/cosmos3-edge-libero10}

Task	Instruction	Success
0	put both the alphabet soup and the tomato sauce in the basket	17/50
1	put both the cream cheese box and the butter in the basket	49/50
2	turn on the stove and put the moka pot on it	49/50
3	put the black bowl in the bottom drawer of the cabinet and close it	49/50
4	put the white mug on the left plate and put the yellow and white mug on the right plate	44/50
5	pick up the book and place it in the back compartment of the caddy	46/50
6	put the white mug on the plate and put the chocolate pudding to the right of the plate	39/50
7	put both the alphabet soup and the cream cheese box in the basket	46/50
8	put both moka pots on the stove	25/50
9	put the yellow and white mug in the microwave and close it	43/50

Notes for running outside the container

torchcodec needs NPP (nvidia/cu13/lib on LD_LIBRARY_PATH) and an FFmpeg with an AV1
decoder (dav1d) for LIBERO's AV1 videos; opencv-python's bundled FFmpeg lacks it.
For the LIBERO sim venv, egl-probe (pulled in by robomimic) fails to build with CMake 4;
CMAKE_POLICY_VERSION_MINIMUM=3.5 fixes it.