ManiSkill2

Last updated: Oct 7, 2026

On this page

Dimensions Data composition Metrics Leaderboard How to run

ManiSkill2 (ICLR 2023, UC San Diego Hao Su Lab with Tsinghua) is the unified Gym generation of the ManiSkill line: 20 task families, 2,000+ object models, and 4M+ demonstration frames on SAPIEN. Families cover stationary and mobile bases, single and dual arms, and rigid plus soft bodies. The same interface exposes point-cloud, RGB-D, and privileged-state observations and several controllers. Official eval reports seen vs unseen assets and configurations. The 2022 challenge used a cloud scorer; that board is not ingested here.

This card is not a rename of ManiSkill. That hub page tracks the later GPU-parallel maniskill.ai importer. Use this page when the export is the ICLR 2023 / ManiSkill2 protocol.

Paper: ManiSkill2: A Unified Benchmark for Generalizable Manipulation Skills. Project: maniskill2.github.io. Code: haosulab/ManiSkill2 (Apache-2.0). Later trees live on haosulab/ManiSkill.

WorldFoundry is official-results only. You run SAPIEN / ManiSkill2 upstream, then official-validation normalizes JSON/CSV rollout summaries. No in-tree simulator, no official policy.

Official paper teaser: a unified suite of stationary/mobile, single/dual-arm, rigid/soft-body tasks.

Dimensions

Stationary / mobile base
KindAxis
Task families on a fixed mount or a mobile base.
Single / dual arm
KindAxis
Single-arm and dual-arm manipulation families.
Rigid / soft body
KindAxis
Rigid objects plus soft-body (cloth / fluid-like) families in SAPIEN.
Observation mode
KindAxis
Point cloud, RGB-D, or privileged state — same Gym interface.
Seen / unseen assets
KindSplit
Official protocol evaluates both seen and held-out object assets and configurations.

Data composition

Size
ICLR 2023 SAPIEN suite: 20 task families, 2,000+ object models, 4M+ demonstration frames. Gym API with multiple controllers. Distinct from the later GPU-parallel maniskill hub card.
Sources
github.com
Splits
seen · unseen
Modalities
vision · action · state
License
apache-2.0
Notes
WorldFoundry imports official rollout summaries. SAPIEN execution and the 2022 cloud eval stay upstream.

Metrics

Metric

success_rate
DirectionHigher is better
Primary episode success on the declared task family and observation mode.
PrimarySuccess Rate
task_success
DirectionHigher is better
Per-family success among the 20 official task families.
episode_success
DirectionHigher is better
Per-episode success indicator when the export lists individual rollouts.
reward
DirectionHigher is better
Mean episode return if the official dump includes dense reward.

Leaderboard

Results from the paper (not a live board; not WorldFoundry-run scores). Official visual-RL baselines on PickCube, not a multi-lab SOTA board. ManiSkill2: A Unified Benchmark for Generalizable Manipulation Skills (2023).

Model	PickCube success (demo-augmented visual RL)
DAPG / PPO, point cloud	0.94
DAPG / PPO, RGBD	0.91

How to run

worldfoundry-eval zoo benchmark-run \
  --benchmark-id maniskill2 \
  --mode official-validation \
  --official-results-path <official_results.json> \
  --output-dir tmp/maniskill2/official-validation \
  --json

Optional WORLDFOUNDRY_MANISKILL2_ROOT points at a local ManiSkill2 / SAPIEN root. Default import path: WORLDFOUNDRY_MANISKILL2_RESULTS_PATH. Manifest: worldfoundry/data/benchmarks/tasks/external/maniskill2.yaml. Outputs: scorecard.json (primary success_rate), raw_results.jsonl, and evaluation/summary.json.

← Benchmark Hub