MultiDojo²Diverse robot teams
Two-, three-, four-, and eight-arm tasks span homogeneous and heterogeneous teams, with explicit roles and instructions.
THE WORLD’S FIRST SELF-EVOLVING HETEROGENEOUS MULTI-ROBOT LEADERBOARD
Teaching robots
to collaborate.
The World's First Self-Evolving Leaderboard for Heterogeneous Multi-Robot Collaboration.
Multi-view training, randomized evaluation, and failure diagnosis guide the next round of learning.
From fan assembly to parallel manipulation: collaboration across homogeneous, heterogeneous, and humanoid robot teams.
20+tasks
Homogeneous & heterogeneous teams
2,000+trajectories
Reviewed training demonstrations
RSI
Diagnose failures. Guide the next round.
Multi-view
Every arm’s first-person + third-person views
01 / THE BENCHMARK
One goal. Many ways to collaborate.
From two-arm handovers to eight-arm parallel tasks,
evaluate the outcome and the collaboration behind it.
Reproducible variation
Vary object positions, destinations, appearance, and initial states within task constraints. Evaluate collaboration against consistent physical criteria.
Scored with physical evidence
Task-specific checks verify grasping, placement, support, stability, and collision constraints. Execution evidence supports reproducible scoring and RSI diagnosis.
Collaboration through everyday manipulation.
Explore all 20+ tasks ↗Loading task demonstrations…
Demonstrations on natural tabletops. Scene variations preserve task instructions and follow a shared scoring protocol.
02 / EVIDENCE-DRIVEN IMPROVEMENT
Understand the failure.
Know what to learn next.
RSI identifies opportunities for improvement
and turns diagnosis into specific training scenarios.
- 01
Train
Aligned observations and actions
Synchronized views
Multi-arm, multi-view data - 02
Evaluate
Randomized environments and initial states
Shared evaluation criteria
Consistent physical validation - 03
Diagnose
RSI assessment
Locate coordination failures
Failure analysis and recommendations - 04
Targeted practice
Specify objects, positions, and roles
Actionable training tasks
Export candidate training scenarios - 05
Retrain & retest
Train on identified weaknesses
Continuous improvement
Retest in independent scenes
RSI / REPLAY → EXPLANATION → CURRICULUM
Feedback that defines the next scenarios.
Link replay timestamps to robot actions and object states to identify the operation stages that need improvement. Recommendations specify tasks, position ranges, motion constraints, and acceptance criteria for candidate scenarios.
Candidate scenes enter targeted training after simulation and data validation. Improvements are then measured under the same evaluation protocol.
03 / START EXPLORING
Trace every step, from data to results.
Multi-view training data
20+ tasks and 2,000+ trajectories, with synchronized first-person views for every arm, third-person views, native observations, and actions. Training data is being progressively released on Hugging Face.
Download training data on Hugging Face ↗Explore data formats and usage ↗Tasks, replays, and results
Randomized trials, task scores, and execution replays follow a shared protocol for policy comparison, reproducible review, and continued improvement.
Open the evaluation workspace ↗Research and usage guide
Task definitions, data formats, evaluation protocols, and the RSI workflow provide a clear starting point for collaboration research and reproducible experiments.
Explore the project ↗Research and usage guide ↗View on GitHub ↗Research directions
MultiDojo² connects multi-view demonstrations, randomized evaluation, and failure-driven training for homogeneous and heterogeneous teams. Research directions include handovers, spatial coordination, timing, and translating execution evidence into specific training tasks.