msolaima/robolab-cosmos3-edge-droid

Cosmos3-Edge-Policy-DROID rollouts in RoboLab

Simulated rollouts of NVIDIA's Cosmos3-Edge-Policy-DROID (4B) on the RoboLab benchmark: a Franka arm, three cameras, and 120 tabletop manipulation tasks. Every episode carries a success label and RoboLab's event log. The data is intended for offline policy evaluation (OPE) and for studying how and why this policy fails.

Status: 91 of 120 tasks, 455 episodes, 113 successes (25%). The remaining 29 tasks will be added to this repo.

Contents

One folder per task (<TaskName>/), with 5 trajectories per task:

File	Contents
traj_0_env<i>.npz	Per control step (T steps at 15 Hz): left_image, right_image, wrist_image (T×224×224×3 uint8), joint_position (T×7), gripper_position (T×1), action (T×8: 7 joint position targets + gripper, binarized), new_chunk (T, true on steps where a new action chunk was fetched from the policy), instruction
log_0_env<i>.json	success, final_step, and RoboLab's event log with step numbers (OBJECT_GRABBED_SUCCESS, TARGET_OBJECT_DROPPED, GRIPPER_HIT_TABLE, WRONG_OBJECT_GRABBED_FAILURE, ...)
run_0.hdf5	RoboLab's per-step record for all 5 trajectories (data/demo_<i>/): joint positions and velocities (7 arm + 6 gripper joints), end-effector pose and velocity, object poses and velocities, object bounding boxes, actions, initial state, subtask status
env_cfg.json	Full RoboLab environment config for the task
videos/episode_0_env<i>_{success,fail}.mp4	Left, wrist and right cameras side by side at 15 fps

At the top level:

File	Contents
results.csv	One row per episode: task, instruction, RoboLab difficulty and attributes, success, length, grasp/drop counts, video path
episode_results.jsonl	RoboLab's own per-episode results (metrics and events)

Load one trajectory:

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download("msolaima/robolab-cosmos3-edge-droid", "BananaInBowlTask/traj_0_env0.npz", repo_type="dataset")
d = np.load(path)
print(d["instruction"], d["wrist_image"].shape, d["action"].shape)

How it was collected

Policy	Cosmos3-Edge-Policy-DROID, served by cosmos-framework 7cfaf78: guidance 3.0 at timesteps 960-1001, 4 denoising steps, 32-action chunks, joint-position actions
Simulator	RoboLab ad45d4f, Isaac Sim 5.0 (Isaac Lab 2.2.0), headless
Tasks	The 120 RoboLab benchmark tasks with their default instruction. Each episode ends on success or at the task's time limit (20-300 s, i.e. 300-4500 steps at 15 Hz).
Trajectories	5 per task, run as 5 parallel environments. All 5 start from the same scene, because RoboLab's benchmark tasks reset to a fixed initial state, so they differ only through the policy's sampling.
Images	Rendered at RoboLab's camera resolution, then stored at 224×224

The collection code (pipeline.sh, collect_robolab_cosmos3.py) lives in data_collection/cosmos3_robolab/ of haowwang/policy_eval_2 (branch msolaima).

Results so far

RoboLab difficulty	Successes
simple	72 / 245 (29%)
moderate	27 / 135 (20%)
complex	14 / 75 (19%)
all	113 / 455 (25%)

11 tasks succeeded in all 5 trajectories, and 56 failed in all 5. The most common failure is grasping the object and dropping it, over and over, until the time limit.

Notes

• Isaac Sim 5.0 logs a Warp CUDA error ... cuDeviceGetUuid warning with the CUDA 13 driver used here. The simulation still ran and finished every episode.
• The data was generated with Cosmos3-Edge-Policy-DROID (OpenMDW 1.1) in RoboLab (Apache 2.0). The images show RoboLab's 3D assets; see the RoboLab repository for their terms.