Contribute

How to grow the benchmark

Anyone interested in imitative ability (goal-directed imitation, video-conditioned policies, cross-embodiment skill transfer, intent inference, affordance adaptation, ...) is welcome to participate. The project grows through four connected contribution paths, all gated by the same design templates and evaluation contracts. Contributions should include enough information for maintainers to reproduce and evaluate them:

Pairing service

Human demonstration data is only useful for imitation when it is paired with robot trajectories. For accepted community-contributed tasks, we may collect human and robot pairing demonstrations and merge the paired data into IG-10K-format datasets under the shared LeRobot format.

Real-device paired data collection

For an accepted task, contributors can provide the human demonstration and task protocol while the project team determines whether matching robot demonstrations can be collected. When available, paired episodes can be released alongside simulated data under the IG-10K / LeRobot contract.

See the data collection guide for preparation details and the Task Gallery for published human, simulated, and real-robot rollouts.

Submit a model or simulation environment

We accept more than standalone checkpoints. To request a real-world evaluation, submit a reproducible bundle through a pull request. Include the model checkpoint or source, inference API, environment dependencies, license, task protocol, reset conditions, task-specific evaluation standard, and commands needed to reproduce inference. We first reproduce the simulation result before scheduling hardware time.

Real-world deployment and evaluation

After the simulation package is reproducible and passes review, contributors can request deployment on the real-robot platform. The team runs the agreed evaluation protocol and records rollout videos and evaluation conditions.

Hardware runs are scheduled only after the package is reproducible. Approved real-world rollouts can be added to the Imitator Arena for blind A/B comparison against the same human reference.

Publish results to Imitator Arena

Once a real-world evaluation is complete, we can add the approved rollouts to Imitator Arena. The Arena presents a human reference and anonymous model rollouts for blind A/B human evaluation, turning a private hardware run into a public, community-readable comparison.

Contribution workflow

1. Follow the design templates: examples/baselines/_template/ for models, mani_skill/envs/tasks/_template/ for L0–L3 tasks, and mani_skill/agents/robots/_template/ for robots.
2. Open a pull request with your task, model, robot, embodiment, or asset contribution.
3. Maintainers review the contribution against the templates (data contract, environment contract, L0–L3 semantics, motion-planning solution).
4. Once accepted, the contribution is merged, registered, and, where applicable, considered for human/robot pairing and release.
5. For real-world evaluation, provide the reproducible model or source, API, dependencies, license, task protocol, reset conditions, evaluation standard, and simulator results.
6. After review, request deployment on the real platform; hardware runs are scheduled only after the package is reproducible.
7. Approved real-world rollouts are published in Imitator Arena for blind A/B community evaluation and linked from the Task Gallery.

Active ManiSkill sync

We keep the simulation framework actively updated with the upstream ManiSkill repository so the community can use the latest simulation features. To keep the ecosystem coherent, all contributions must follow the shared ManiSkill design templates — the robot agent template, the environment interface, and the motion-planning interface are the same ones the upstream project documents. If you build on a new ManiSkill feature, please keep the common interfaces intact.

Join the community

Questions about the benchmark, data, or a potential contribution are welcome. Use whichever community channel is most convenient:

Discord

Join the project community for announcements, technical discussion, and collaboration.

Join on Discord

WeChat community

For Chinese-language discussion, scan the QR code to join the project group.

Scan to join the WeChat group

For direct technical inquiries, contact Xunzhe Zhou.

Citation

@unpublished{imitatorgame2026,
  title={The Imitator Game: Benchmarking Robot Imitative Ability Beyond Action Prediction},
  author={Zhou, Xunzhe and Cai, Yiyang and Wang, Fengyi and Ju, Ran and Ren, Hanxiang and Liu, Ruizhe and Zhang, Yu and Luo, Qian and Chen, Feng and Zhou, Pei and Ma, Yi and Yang, Yanchao},
  url={https://github.com/imitator-game/The-Imitator-Game},
  year={2026}
}