RoboLab-EgoX#

Policy rollouts recorded on NVIDIA’s RoboLab manipulation benchmark, as native .mcap episodes.

Each take carries three synchronized camera views with a matching 16-bit depth stream, per-camera intrinsics, joint positions, actions, end-effector pose, and the task instruction. Takes keep the benchmark’s own success label, so failed rollouts sit alongside successful ones: 632 of the 4,000 takes succeeded, spanning all 28 tasks and 99 background scenes.

In 153 of the 4,000 takes the camera streams carry 80 frames while depth and telemetry carry 81, so align streams by timestamp rather than by index.

Details

Example usage

1import fiftyone as fo
2import fiftyone.zoo as foz
3
4dataset = foz.load_zoo_dataset("robolab")
5
6# Rollouts the policy got right
7view = dataset.match({"success": True})
8
9session = fo.launch_app(dataset)
robolab