RoboCasa v0.2: 7,200 Robot Task Demos with Depth and RGB

Per-step environment depth and RGB sidecar images rendered from robot demonstration data. The dataset contains approximately 7,200 demos across 24 kitchen tasks, filtered to include only successful task completions. It was created by SeonghuJeon and last updated on Hugging Face in May 2026.

Use Cases

• Training robot manipulation policies based on successful demonstration sequences.
• Developing models for depth and RGB image fusion in simulated environments.
• Benchmarking imitation learning algorithms on a set of 24 atomic kitchen tasks.
• Analyzing the geometry of successful task executions from multi-camera perspectives.

Strengths

• Contains 300 success-audited demonstrations per task, providing a consistent quality filter.
• Covers 24 distinct atomic kitchen tasks for varied manipulation scenarios.
• Images are rendered at a native 256-pixel resolution.
• Data is structured as per-camera contiguous .npy memmaps for efficient access.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is derived from a specific simulation environment (RoboCasa), which may limit real-world applicability.

Provenance

Source

Rendered from the original RoboCasa v0.2 mg_im Generated-300 HDF5 demos.

Collection Method

Demos were generated in simulation and filtered to those whose final state matches a task-success predicate.

Freshness

Last updated 2026-05-20 05:40:19; freshness should be verified.