ManiSkill

Last updated: Oct 7, 2026

On this page

Dimensions Data composition Metrics Leaderboard How to run

ManiSkill is UC San Diego Hao Su Lab's GPU-parallel manipulation benchmark on SAPIEN: dozens of skill families, multiple robot embodiments, demonstration datasets, and massively parallel rollouts for RL and imitation. The paper is ManiSkill: Generalizable Manipulation Skill Benchmark with Large-Scale Demonstrations. Project: maniskill.ai. Code: haosulab/ManiSkill (Apache-2.0). Later ManiSkill2 / ManiSkill3 releases keep the same lab and engine; this card is the ManiSkill importer, not a silent rename of those pages.

WorldFoundry's path is official-results only. You run the upstream tooling, then official-validation normalizes JSON/CSV rollout summaries into the unified scorecard. Simulator execution, policy loading, and data collection stay explicit and upstream. The repo mirror is cache/vla_va_wam/repos/haosulab--ManiSkill.

Official paper teaser: OpenCabinet, PushChair, and MoveBucket across many articulated objects.

Dimensions

Rigid-body manipulation
KindCategory
Pick, stack, insert, and related rigid-object families on SAPIEN.
Soft / articulated objects
KindCategory
Deformable and articulated assets where contact and joint limits matter.
Robot embodiments
KindAxis
Multiple arms and grippers; the official suite is GPU-parallelized on SAPIEN.
Demonstrations vs RL
KindTrack
Large demonstration sets for imitation, plus massively parallel RL rollouts.

Data composition

Size
GPU-parallel SAPIEN benchmark from UCSD Hao Su Lab: dozens of manipulation families, multiple embodiments, and demonstration datasets. WorldFoundry does not run the simulator; it normalizes official rollout exports.
Sources
github.com
Modalities
language · vision · action · state
License
apache-2.0
Notes
Simulator execution, policy loading, and rollout collection stay upstream. This page is an official-results importer.

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.
reward
DirectionHigher is better
Mean episode reward / return.
normalized_return
DirectionHigher is better
Normalized episode return.

Leaderboard

Results from the paper (not a live board; not WorldFoundry-run scores). Same official paper as ManiSkill2. From-scratch visual RL on PickCube. ManiSkill2: A Unified Benchmark for Generalizable Manipulation Skills (2023).

Model	PickCube success (from-scratch visual RL)
PPO, point cloud	0.22
PPO, RGBD	0.01

How to run

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

Default import path when you omit --official-results-path is WORLDFOUNDRY_MANISKILL_RESULTS_PATH. Task manifest: worldfoundry/data/benchmarks/tasks/external/maniskill.yaml. Outputs: scorecard.json (primary success_rate) and raw_results.jsonl.

← Benchmark Hub