UB Robotics · NVIDIA Cosmos Codefest 2026

Cosmos3-Edge on LIBERO-10

A post-training recipe that turns the public nvidia/Cosmos3-Edge world model into a robot-arm policy for the LIBERO-10 benchmark: ten long-horizon kitchen and tabletop tasks with a Franka arm in simulation. The recipe is proposed upstream in NVIDIA/cosmos-framework PR #278. The trained model is public: ubr-physical-ai/cosmos3-edge-libero10.

Closed-loop success, 500 episodes

81.4%

Successes: 407 / 500
Checkpoint: iter 2000
Trials per task: 50
Tasks above 85 %: 7 of 10
Model: download (public)

The policy at work

One episode per LIBERO-10 task with the final checkpoint, agent-view camera, instruction on top. All ten of these succeeded; the success rate above comes from the separate 500-episode run.

Per-task results

#	Instruction	Success

Two tasks pull the average down: putting both the alphabet soup and the tomato sauce in the basket (17/50), and placing both moka pots on the stove (25/50). Both need two pick-and-place moves of similar objects in one episode.

Without fine-tuning

The same evaluation run on the public nvidia/Cosmos3-Edge checkpoint with no LIBERO training: same policy server, sampler, cameras and simulator settings. NVIDIA has not published a LIBERO number for Edge; this is our measurement. The released checkpoint has a trained action head for some robots, but its LIBERO slot is still at its initial values (never trained by NVIDIA), so this measures an untrained LIBERO output: a floor, not zero-shot skill.

Vanilla Cosmos3-Edge, 500 episodes

0.0%

Successes: 0 / 500 (0 of 50 on every task)
Recorded run: 0 / 10 (every episode hit the 520-step limit)
Fine-tuned: 407 / 500 = 81.4 %

Same task and start state, one episode per task. Left: fine-tuned with this recipe. Right: the vanilla checkpoint, which never completes a task.

How it compares

Published LIBERO-10 (LIBERO-Long) success rates of other fine-tuned policies. Every row except the vanilla Cosmos3-Edge is a model post-trained on LIBERO by its authors; the vanilla row has an untrained LIBERO action head. Inputs, training data and trial counts differ, so this is context, not a controlled comparison.

Model	Size	LIBERO-10		Source

Sizes are total parameters as published by each project. Diffusion Policy is trained from scratch and its size is not given in the source table.
Cosmos Policy trains on all four LIBERO suites with wrist camera and robot state, 40,000 steps on 64 H100s, and reports 3 seeds of 500 trials. GR00T N1.6 reports 200 episodes. Our run trains on LIBERO-10 only, 2,000 iterations on 4 GPUs, 500 trials.
π0-FAST's 60.2 % is the widely cited figure; an openpi issue reports about 85 % with a configuration change.

Training loss

Loss logged every iteration (faint) and its 25-iteration moving average (solid). The learning rate warms up over the first 500 iterations.

The run

Base modelnvidia/Cosmos3-Edge (4B, Mixture-of-Transformers)

DataLIBERO-10, LeRobot v3

Hardware4 × RTX PRO 6000 Blackwell 96 GB

Wall time3 d 4 h, 2,000 iterations

Global batch2,048 (128 × 4 GPUs × grad-accum 4)

Optimizerlr 5e-5, 500 warm-up steps

Precisionbfloat16, FSDP shard 4

Actions10-D, frame-wise relative, rot6d

Weightsubr-physical-ai/cosmos3-edge-libero10 (iteration 2000, public)

The upstream recipe runs on two 8-GPU nodes (HSDP 2 × 8, grad-accum 1). We kept the same experiment, learning-rate schedule and global batch and changed only the parallelism to fit four GPUs.

How it was evaluated

Policy server: action_policy_server_libero on the final checkpoint, quantile-rot6d action normalization with the bundled LIBERO stats, 30 sampling steps, 20 fps.
Simulator client: closed_loop_eval.py, agent-view + wrist cameras at 256 px, 50 trials per task, 8 parallel environments, seed 0.
Success is the LIBERO task predicate at episode end (520-step limit).

python cosmos_framework/simulation/libero/closed_loop_eval.py \
  --server_url http://localhost:8000 --task_suite libero_10 \
  --num_trials_per_task 50 --num_envs 8 \
  --camera agentview,wrist --image_size 256 \
  --action_space frame_wise_relative --rotation_space 6d --action_dim 10

Setup note: the LIBERO simulator environment pulls in egl-probe, which does not build with CMake 4; setting CMAKE_POLICY_VERSION_MINIMUM=3.5 fixes it.