RoboCasa

Last updated: Oct 7, 2026

On this page

Dimensions Data composition Metrics Leaderboard How to run

RoboCasa is UT Austin and NVIDIA's large-scale kitchen simulation for generalist robots: 120 realistic kitchen layouts, 2,500+ 3D assets (many generated), and about 100 atomic skills plus composite long-horizon activities, with large demonstration sets. The paper is RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots. Project: robocasa.ai. Code: robocasa/robocasa. GitHub license metadata in this catalog is other:noassertion_github_api — read the upstream terms before reuse.

WorldFoundry imports official rollout exports. Scene assets and long-horizon policy execution stay upstream. Repo mirror: cache/vla_va_wam/repos/robocasa--robocasa. Models such as Being-H0.5 publish a RoboCasa specialist line; that is not a score on this page.

Official paper teaser: robots in diverse simulated kitchens.

Dimensions

Atomic kitchen skills
KindSuite
About 100 atomic manipulation skills in kitchen scenes.
Composite long-horizon
KindSuite
Composite tasks that chain atomic skills into longer kitchen activities.
Kitchen scenes
KindAxis
120 realistic kitchen layouts and 2,500+ 3D assets (many generative).

Data composition

Size
UT Austin & NVIDIA kitchen simulation: 120 scenes, 2,500+ assets, 100 atomic plus composite long-horizon tasks, with large demonstration sets for generalist policies.
Sources
github.com
Splits
kitchen · long_horizon
Modalities
language · vision · action · state
License
other:noassertion_github_api
Notes
RoboCasa uses simulator assets and generated policy rollouts rather than a single public HF benchmark table.

Metrics

Metric

success_rate
DirectionHigher is better
Primary episode success rate over evaluated rollouts.
PrimarySuccess Rate
task_success
DirectionHigher is better
Per-task success aggregate.
episode_success
DirectionHigher is better
Per-episode success indicator aggregate.
reward
DirectionHigher is better
Mean episode reward / return.

Leaderboard

Results from the paper (not a live board; not WorldFoundry-run scores). Same BC policy trained on different data mixes from the paper. RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots (2024).

Model	Average success (24 atomic tasks)
Generated-3000	0.476
Generated-300	0.350
Human-50	0.288
Generated-100	0.263

How to run

worldfoundry-eval zoo benchmark-run \
  --benchmark-id robocasa \
  --mode official-validation \
  --official-results-path <official_results.json> \
  --output-dir tmp/robocasa/official-validation \
  --json

Default import path: WORLDFOUNDRY_ROBOCASA_RESULTS_PATH. Manifest: worldfoundry/data/benchmarks/tasks/external/robocasa.yaml. Declare atomic vs composite (and the scene split) with the export. Outputs: scorecard.json (primary success_rate) and raw_results.jsonl.

← Benchmark Hub