SIMPLE: Simulation-Based Policy Learning and Evaluation for Humanoid Loco-manipulation

Abstract: Humanoid foundation models are advancing faster than we can evaluate them. While real-world testing is expensive and difficult to reproduce, existing simulation benchmarks focus primarily on table-top or wheeled robots. A scalable and reproducible benchmark for whole-body humanoid loco-manipulation remains an open problem. To this end, we present SIMPLE, a unified simulation testbed for humanoid policy learning and evaluation. SIMPLE couples the accurate contact-rich dynamics of MuJoCo with the photorealistic rendering of IsaacSim. It provides a large-scale environment comprising 60 diverse whole-body tasks, 50 indoor scenes, and over 1,000 object assets. To facilitate scalable data collection, the framework integrates two data generation pipelines: automated trajectory generation via motion planning and a low-latency VR teleoperation interface. We further integrate and benchmark mainstream humanoid policies at scale in SIMPLE, including lightweight imitation networks, large vision-language-action (VLA) models, and recent world action models (WAMs). Our experiments reveal a strong correlation between policy performance in simulation and the real world. Furthermore, we demonstrate that policies trained on data collected in SIMPLE can be transferred zero-shot to physical humanoid robots under similar settings, providing a robust and reproducible foundation for humanoid robotics research.

• The paper introduces a dual-simulator architecture that couples high-frequency physics (MuJoCo) with photorealistic rendering (Isaac Sim) to enhance evaluation fidelity.
• It presents scalable data generation pipelines using automated motion planning and VR teleoperation for robust and efficient policy training.
• Policies trained using SIMPLE demonstrate strong zero-shot sim-to-real transfer, with simulation rankings closely mirroring real-world performance.

SIMPLE: Unified Simulation-Based Policy Learning and Evaluation for Humanoid Loco-manipulation

Motivation and Related Work

The evaluation and training of generalist humanoid robotics policies are fundamentally hampered by a lack of scalable, reproducible benchmarks. Existing platforms have accelerated progress in tabletop or wheeled robots via standardized simulation environments, but these are not directly extensible to whole-body humanoid loco-manipulation, which requires coordinated balance, locomotion, and complex object interaction. Real-world evaluation remains the dominant paradigm but is labor-intensive, unreproducible, and expensive, while crowd-sourced rankings and LLM-automated resets are limited by scalability and bias. Previous simulation benchmarks—RLBench, LIBERO, RoboCasa, Behavior-1K, etc.—lack the physics fidelity or diverse asset integration necessary for whole-body humanoid policy learning and evaluation. Simulation-based frameworks such as AMO and SONIC have demonstrated robust locomotion but have not coupled this with visually rich, contact-dense manipulation settings.

Framework Architecture

SIMPLE introduces a dual-simulator architecture that decouples physics simulation and photorealistic rendering to maximize both contact fidelity and visual diversity. MuJoCo serves as the physics backend, running at 500 Hz to simulate high-frequency contact interactions and whole-body balancing. Isaac Sim provides visually rich ray-traced RGB observations from synchronized physical states at 50 Hz. The environment is wrapped in a standard OpenAI Gym interface to facilitate integration with RL pipelines.

SIMPLE implements a modular whole-body control scheme: high-level policies (VLAs, WAMs) output kinematic targets and navigation commands, which are tracked by integrated lower-body controllers (AMO and SONIC). The asset library comprises over 1,000 objects and 50 indoor scenes, processed via CoACD for accurate collision geometry and converted to USD format for high-resolution rendering. Tasks span rigid pick-and-place, articulated manipulation, non-prehensile interaction, and whole-body loco-manipulation across spatially diverse layouts.

Scalable Data Generation Pipelines

SIMPLE standardizes demonstration data collection through two pipelines:

• Automated Motion Planning: CuRobo synthesizes trajectories based on task decomposition and offline grasp pose generation (BoDex), enabling fully automated script-driven demonstrations for high-throughput data generation without operator intervention.
• VR Teleoperation: Low-latency teleoperation streams egocentric stereo views to a PICO XR headset, retargeting operator hand motions to robot via inverse kinematics. Whole-body tracking controllers autonomously maintain balance and locomotion, minimizing operator cognitive load.

Both pipelines are designed to produce high-quality demonstration trajectories, which are subsequently replayed with extensive domain randomization in Isaac Sim for robust policy learning under varying visual and spatial conditions.

Benchmarking Protocol and Experimental Evaluations

SIMPLE defines three levels of domain randomization for systematic policy evaluation:

• Level 0: Random distractor objects and backgrounds
• Level 1: Additional material and lighting randomization
• Level 2: Spatial randomization of object and robot initial poses

Mainstream policy architectures (Yo, To.5, GR00T-N1.6, DreamZero, EgoVLA, H-RDT, ACT, DP) are benchmarked across six representative tasks and all three DR levels. Numerical results highlight several critical findings:

• Yo achieves consistently high success rates across all tasks and DR levels, outperforming baselines in most settings.
• DreamZero and To.5 exhibit strong generalization to visually randomized settings, with a notable performance drop in tasks requiring precise base movement.
• ACT demonstrates strong data efficiency, attributed to the low noise in simulator-generated teleoperation data.

The simulation rankings closely mirror real-world policy performance, evidencing a strong correlation and validating the use of SIMPLE as a reliable proxy for real-world evaluation.

Data Collection Efficiency and Ablative Analysis

Comparative throughput analyses reveal:

• Simulated teleoperation achieves the highest demonstrations/hr, eliminating hardware reset overhead and operator fatigue.
• Motion planning is less efficient but scales autonomously, while real-world teleoperation provides the most natural motion profiles at moderate throughput.
• Data collected in simulation enables infinite scaling and diverse scene composition via offline rendering.

Ablation studies show that mixed-domain randomization results in improved cross-level generalization. Scaling teleoperation data substantially boosts policy performance. Teleoperation-generated data yields greater effectiveness compared to scripted motion planning, especially for dexterous and contact-rich tasks.

Zero-Shot Sim-to-Real Transfer

SIMPLE-trained policies exhibit robust zero-shot transfer to physical humanoid robots without downstream real-world fine-tuning. Evaluation on a subset of tasks under matched environmental settings demonstrates direct applicability, highlighting that the combination of contact-accurate physics and photorealistic rendering is sufficient to bridge the sim-to-real gap for whole-body loco-manipulation. This is a critical empirical claim substantiated with strong numerical results.

System Implementation Details

SIMPLE's internal architecture comprises:

• BaseDualSimEnv: Manages MuJoCo and Isaac Sim instances, task configuration, and Gym integration.
• Simulator Classes: Abstracted backend implementations for physics and rendering with message-passing synchronization.
• Task and Robot Classes: Define configuration, asset paths, domain randomization, and fine-grained control attributes.
• Agent Hierarchy: Unified get_action interface across motion planning, teleoperation, VLA/WAM inference, supporting both local and remote server processes.
• DomainRandomizer: Extensive randomization registry for materials, spatial layouts, lighting, and distractor distributions.

Offline asset preprocessing via CoACD and BoDex generates physics-ready collision geometries and cached dexterous grasps. Integrated locomotion controllers (AMO, SONIC) manage high-dimensional whole-body tracking at 500 Hz for physically plausible demonstration collection and policy execution.

Limitations and Practical Implications

Two primary limitations are outlined:

• Rendering throughput: Photorealistic ray-tracing is computationally intensive (~4 FPS), constraining large-scale data generation.
• Rigid-body assumption: Deformable and soft-body objects are not currently supported, limiting the simulation's coverage for certain manipulation categories.

Despite these, SIMPLE represents a reproducible, extensible foundation for benchmarking and training generalist humanoid policies in simulation. Practically, it enables rapid, fair, and scalable evaluation, reducing reliance on expensive real-world testing. Theoretically, it sets the stage for systematic analysis of policy architectures and sim-to-real transfer. The open-source release will facilitate community-wide adoption, extension, and standardization.

Conclusion

SIMPLE delivers a full-stack, scalable simulation framework that addresses fragmentation in humanoid robotics benchmarking. The coupling of high-fidelity contact physics and photorealistic rendering supports both rigorous policy evaluation and robust data generation at scale. Extensive experiments demonstrate that performance in SIMPLE closely tracks real-world outcomes, and policies trained on SIMPLE data exhibit effective zero-shot sim-to-real transfer. These capabilities set a new standard for reproducible evaluation and accelerated research in humanoid loco-manipulation, with broad implications for the development of generalist robotic foundation models (2606.08278).