Robots Atlas>ROBOTS ATLAS
Evaluation

RoboDojo

2026ActivePublished: 23 September 2026Updated: 23 September 2026Published
Key innovation
Unifies evaluation of generalist robot manipulation policies by combining simulation and reproducible real-world testing (RealEval) in a single benchmark, instead of sim-only or real-only evaluation.
Category
Evaluation
Abstraction level
System
Operation level
Robot controlSystem
Use cases
Evaluation of generalist robot manipulation policiesComparable sim-to-real testingEmbodied AI and manipulation robotics researchPublic leaderboard for robot policiesReproducible real-world evaluation (RealEval)

How it works

The benchmark provides a simulator client and a suite of evaluation tasks running on Isaac Sim, with heterogeneous parallel simulation of different tasks and scenes for fast, scalable feedback. Scenes are built from physically grounded, configuration-driven assets (rigid, articulated, and deformable objects), with seed-controlled layouts for reproducibility. The RoboDojo-RealEval layer runs analogous tests on physical robots (Piper X, Piper, ARX X5) with remote cloud access, standardized hardware, and scene reset. Policies are attached through the shared XPolicyLab interface, and a single command (summarize) aggregates results into a leaderboard table. In the current release RoboDojo is eval-only — it ships the simulator, tasks, asset/config validation, and result artifacts, while policy integration and policy servers belong to XPolicyLab.

Problem solved

Prior manipulation benchmarks were typically sim-only or real-only, which made scalable, reproducible, and comparable evaluation of generalist policies difficult. RoboDojo bridges both regimes and standardizes the evaluation protocol so results are reproducible and comparable across models.

Components

Simulation task suite (Isaac Sim)Simulation part of the benchmark

42 Isaac Sim tasks grouped into five capability dimensions (generalization, memory, precision, long-horizon, open tasks), with heterogeneous parallel simulation.

RoboDojo-RealEvalReproducible real-world evaluation

18 tasks on physical robots (Piper X, Piper, ARX X5) with remote cloud access, standardized hardware, scene reset, an evaluation protocol, and a deployment interface.

XPolicyLabPolicy integration and serving

Companion project unifying many policies behind a single interface, used for both simulation and real-world runs; in the eval-only RoboDojo release, policy integration belongs to XPolicyLab.

LeaderboardResult aggregation and comparison

Public, continuously updated ranking; results aggregated with a single summarize command over seed-controlled layouts for reproducibility.

Implementation

Implementation pitfalls
Eval-only releaseMedium

RoboDojo ships only the simulator, tasks, asset/config validation, and result artifacts; policy integration and policy servers are not part of the repository.

Fix:Use the companion XPolicyLab project to integrate and serve policies.
Isaac Sim / GPU memory requirementsMedium

Simulation relies on Isaac Sim and requires sufficient GPU memory; the documentation lists common issues with installation, assets, and GPU memory.

Fix:Consult the Common Issues section in the documentation and provide a compatible GPU/Isaac Sim environment.

Evolution

Original paper · 2026 · arXiv preprint (2026) · Tianxing Chen
RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies
Tianxing Chen, Hao Dong, Wojciech Matusik, Masayoshi Tomizuka
2026
Paper release and open-source code
Inflection point

RoboDojo was published on arXiv (2607.04434) and its repository was open-sourced; the benchmark ships as eval-only.