AIAny
Icon for item

RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies

Evaluates generalist robot manipulation policies across simulation and real-world settings using 42 sim tasks and 18 real tasks; measures generalization, memory, precision, long-horizon execution and open-vocabulary instruction following, and provides a cloud-accessible real-world evaluation system with XPolicyLab integration and a public leaderboard.

Introduction

RoboDojo tackles a practical gap: benchmarks either scale in simulation or expose real-world challenges, but rarely unify both with reproducible protocols. By combining heterogeneous parallel simulation with a standardized, remotely accessible real-world evaluation system, the paper makes it possible to compare generalist manipulation policies on the same set of capabilities and deployment constraints.

Key Findings
  • Multi-domain coverage: The benchmark pairs 42 simulation tasks with 18 real-world tasks to span complementary manipulation capabilities. This means evaluations capture both scalable metric-driven comparison in sim and the deployment variability of physical systems.
  • Five evaluation dimensions: Generalization, memory, precision, long-horizon execution, and open-vocabulary instruction following are measured explicitly—so evaluators can pinpoint which aspects of policy performance drive failures in transfer or deployment.
  • Reproducible real-world eval: RoboDojo-RealEval offers standardized hardware, scene reset procedures, evaluation protocol, and remote cloud access, enabling repeatable physical tests without each group rebuilding hardware setups.
  • Integration and benchmarking: XPolicyLab lets policies be integrated once and run across sim and real with minimal adaptation; the authors integrated 30 policies and published a public leaderboard to surface current performance gaps.
Who It's For and Trade-offs

Great fit if you need a systematic, comparable evaluation between sim and real for manipulation research—particularly for teams benchmarking generalist policies across multiple capabilities or studying sim-to-real gaps. The benchmark accelerates comparative analysis and reproducibility by providing task suites, standardized real-world rigs, and an evaluation pipeline.

Look elsewhere if your focus is on highly specialized hardware, non-manipulation robotics (e.g., locomotion-only), or ultra-low-latency on-device deployment: RoboDojo’s real-world arm is standardized and queue-based cloud access may not match bespoke lab setups. Also, while the suite is broad, adding new hardware or radically different task modalities will require extra engineering to adopt the provided RealEval rig.

Overall, RoboDojo is positioned as a practical middle ground for researchers who want both scalable simulation studies and reproducible physical validation without rebuilding end-to-end infrastructure.

Information

  • Websitearxiv.org
  • AuthorsTianxing Chen, Yue Chen, Zixuan Li, Junyuan Tang, Kailun Su, Haoran Lu, Weijie Wan, Baijun Chen, Songling Liu, Haowen Yan
  • Published date2026/07/05

Categories

More Items

Hugging Face

Provides 90,000 hours of head-mounted egocentric video paired with synchronized 3D hand pose and an optional 3D full‑body pose add-on, with event-level semantic labels available as a complimentary layer — designed for embodied AI and robotics training at scale.

Enables closed-loop execution for embodied agents by evolving code-based runtime critics and recovery skills online while keeping the base policy frozen. Combines three timescale loops with Z-Infra rollout infrastructure; reports 90.8% on LIBERO-Pro, 93.6% on RoboCasa and an 11.1× inference speedup.

Turns embodied navigation into 2D visual prompting where a vision-language model selects image pixels that are projected to 3D actions; adds selective chain-of-thought, compressed anchor-trajectory memory, and a two-level alignment objective to improve sample and runtime efficiency.