Space Robotics Bench

• Space Robotics Bench is an open-source framework that generates vast, diverse training distributions to overcome data scarcity and high hardware costs in space robotics.
• The system integrates NVIDIA Isaac Sim with specialized modules for asset management, procedural generation, and domain randomization to support scalable reinforcement learning.
• Its modular, composable design supports varied robot embodiments and mission-relevant tasks, enabling effective sim-to-real transfer and robust extraterrestrial autonomy.

Searching arXiv for papers on "Space Robotics Bench" and closely related benchmark frameworks to ground the article. Space Robotics Bench (SRB) is an open-source simulation framework and benchmark suite for robot learning beyond Earth, introduced to address data scarcity, expensive hardware testing, and poor generalization under unknown extraterrestrial conditions in space autonomy research (Orsula et al., 27 Sep 2025). It is built on NVIDIA Isaac Sim and Isaac Lab, but extends them with space-robotics-specific systems for asset management, scenario handling, task logic, procedural generation, domain randomization, simulation parallelization, Gymnasium and ROS 2 integration, and sim-to-real deployment workflow. The framework is positioned not as a mission-specific digital twin or a single verification environment, but as research infrastructure for generating vast and diverse training distributions and for benchmarking algorithms on a common suite of mission-relevant tasks (Orsula et al., 27 Sep 2025).

1. Motivation and benchmark philosophy

The motivation for SRB is that space robotics cannot rely on the same data regime as terrestrial robot learning. Real missions are expensive, rare, and risky, which makes collection of large training datasets on hardware infeasible. The framework is introduced against a background in which existing space simulators are described as mission-specific and tightly scoped to one robot or one environment, often restricted-access or agency-only, optimized for verification and validation of a single system rather than scalable learning, and not designed for massive diversity and throughput of the kind required by modern reinforcement learning (Orsula et al., 27 Sep 2025).

A central premise of SRB is that the sim-to-real gap is not best addressed by seeking a single perfect simulator. Instead, the framework advances the view that one should generate many plausible realities and force the agent to learn invariants across them. This places diversity, procedural variability, and broad training distributions at the center of the benchmark design rather than treating them as auxiliary robustness measures (Orsula et al., 27 Sep 2025).

This orientation distinguishes SRB from benchmark and testbed efforts that emphasize a uniform interface across simulation and hardware, such as BenchBot, or a cross-medium navigation suite such as NavBench. BenchBot formalizes a workflow of selecting a task, platform, and environment, submitting a solution, and receiving evaluation feedback through a common interface, while NavBench standardizes task definitions, observations, rewards, success criteria, training pipeline, and ROS2 deployment across land, water, and space-like navigation platforms (Talbot et al., 2020, El-Hariry et al., 20 May 2025). SRB adopts a related concern for standardization, but applies it to robot learning in extraterrestrial domains with a broader task set and a stronger emphasis on procedural diversity (Orsula et al., 27 Sep 2025).

2. System architecture and software organization

SRB is structured around modularity and extensibility. New assets can be added in USD format plus a declarative Python class, and robots, end-effectors, actuation models, and tasks are all composable. Multi-component systems automatically build a unified action space, so that, for example, a mobile manipulator can combine base, arm, and gripper models. The same task logic can therefore be reused across multiple robots and domains (Orsula et al., 27 Sep 2025).

The architecture adds space-robotics-specific systems on top of Isaac Sim and Isaac Lab for asset management, scenario handling, task logic, procedural generation, domain randomization, simulation parallelization, Gymnasium and ROS 2 integration, and sim-to-real deployment workflow. Accessibility is supported through a Gymnasium API, ROS 2 support, and compatibility with reinforcement learning libraries such as Stable-Baselines3 and skrl (Orsula et al., 27 Sep 2025).

This modular decomposition is significant because space robotics spans wheeled rovers, legged robots, aerial vehicles, spacecraft, manipulators, end-effectors, mobile manipulators, and humanoids. A benchmark that treats these as composable rather than isolated categories can support studies of cross-morphology policy learning and generalization. A plausible implication is that SRB is intended not only for task benchmarking within a single embodiment class, but also for controlled comparison across embodiments under shared task logic (Orsula et al., 27 Sep 2025).

Comparable modularity appears in other benchmark infrastructures, though with different emphases. CoBRA decomposes a benchmark into robot or module set, cost function, and task, with JSON-based descriptions for robots, environments, and tasks, and explicitly supports modular robots and joint optimization over robot assembly, base pose, and trajectory (Mayer et al., 2022). RobotPerf likewise separates workloads across perception, localization, control, and manipulation while benchmarking the robotics computing system through ROS 2 packages and black-box or grey-box instrumentation (Mayoral-Vilches et al., 2023). SRB extends the modular benchmark idea into the specific regime of extraterrestrial robot learning (Orsula et al., 27 Sep 2025).

3. Procedural generation, domain randomization, and parallel simulation

A major contribution of SRB is its on-demand procedural content generation pipeline. The framework requests unique asset variants at runtime, one per parallel environment. A procedural engine built on Blender generates assets using node-based parametric blueprints and can synthesize terrains, craters, rocks and samples, spacecraft, tools, robot embodiments, and end-effectors. The paper gives a lunar terrain example in which generation begins from a flat mesh, applies low-frequency Perlin noise for large-scale topography, adds higher-frequency noise, and uses Voronoi noise to create crater rims and sharper features. Assets are automatically exported, materials are baked into PBR textures, and the results are saved as USD for efficient loading (Orsula et al., 27 Sep 2025).

The importance of this system is its ability to generate hundreds of unique procedural worlds in seconds, creating a near-unlimited training distribution. The framework ties this directly to the needs of reinforcement learning: large-scale data, reduced overfitting, variation in world and embodiment, and improved sim-to-real robustness through exposure to many plausible variants of reality (Orsula et al., 27 Sep 2025).

SRB is also designed for high-throughput parallel simulation. It uses GPU capabilities from Isaac Sim and Isaac Lab, supplemented with Rust extension modules via PyO3 for CPU-bound logic, TorchScript for GPU-side task logic, and a unified command-line interface to control the number of environments and task configuration. Reported throughput illustrates the range of computational regimes: landing can exceed 100k steps/sec with 1024 environments; many locomotion and manipulation tasks plateau around 15k steps/sec with 512 environments; excavation with particle simulation is much slower, around 40 steps/sec (Orsula et al., 27 Sep 2025).

The emphasis on massively parallel simulation places SRB in continuity with other benchmark systems that use simulation not simply for convenience but for structured performance characterization. BenchBot supports sweeps over robot-environment combinations through benchbot_batch to generate more complete performance profiles, and SpaceSense-Bench uses a fully automated Unreal Engine 5 pipeline with synchronized RGB, depth, and LiDAR capture plus automated quality control and export to mainstream formats (Talbot et al., 2020, Wu et al., 10 Mar 2026). SRB differs in that its primary target is large-scale learning rather than only evaluation or dataset generation (Orsula et al., 27 Sep 2025).

4. Domains, robots, actuation models, and benchmark tasks

SRB defines preset space-relevant domains with distinct gravity and visual conditions: Orbit with $g = 0.00$, Asteroid with $g = 0.14 \pm 0.14$, Moon with $g = 1.62 \pm 0.01$, Mars with $g = 3.72 \pm 0.01$, and Earth with $g = 9.81 \pm 0.03$ (Orsula et al., 27 Sep 2025). It includes wheeled rovers, legged robots, aerial robots, spacecraft or landers, manipulators, end-effectors, mobile manipulators, and humanoids, together with modular actuation models including wheeled velocity or joint control, legged or humanoid joint position control, aerial acceleration control, spacecraft thruster and acceleration control, manipulator control via OSC, IK, or joint position control, and end-effector joint control (Orsula et al., 27 Sep 2025).

The benchmark suite is organized into three major task groups.

Group	Tasks
Mobile robotics	landing; rendezvous; orbital_evasion; velocity_tracking; locomotion_velocity_tracking; waypoint_navigation; locomotion_waypoint_navigation; orbital_waypoint_navigation
Fixed-base manipulation	debris_capture; sample_collection; multi_sample_collection; excavation; peg_in_hole; multi_peg_in_hole; screwdriving; solar_panel_assembly
Mobile manipulation	mobile_debris_capture; mobile_excavation

The task design is explicitly mission-relevant and holistic. Rather than isolating narrowly defined subtasks, SRB emphasizes complete behaviors such as full pick-and-place, full trajectory tracking, or multi-stage assembly (Orsula et al., 27 Sep 2025). This places it closer to mission scenario benchmarking than to narrowly scoped control benchmarks.

Several related papers in the broader literature illuminate the relevance of these task categories for space robotics. ReachBot frames sparse-contact locomotion and compliant, tension-dominated mobile manipulation as a scientifically interesting benchmark platform for low-gravity climbing and traversing (Schneider et al., 2021). ReCoBot provides a 7-axis walking space manipulator that can locomote over a satellite mockup through repeated docking and planning events (Scherzinger et al., 2022). Hardware-in-the-loop and air-bearing testbeds for on-orbit servicing similarly foreground rendezvous, docking, capture, coupled dynamics, and manipulation under spacecraft reaction effects (Sah et al., 2023, Korczyk et al., 2022). SRB can be understood as a software benchmark suite that subsumes many of these mission archetypes into a common learning-oriented framework (Orsula et al., 27 Sep 2025).

5. Learning formulation and baseline methodology

SRB tasks are formulated as partially observable Markov decision processes with multi-modal observations. Observations are organized as a dictionary of tensors with four categories: State, Proprioception, Visual, and Commands. State contains privileged simulation information; Proprioception includes internal measurements such as kinematics, IMU, and fuel; Visual comprises RGB, depth, normals, and segmentation; Commands represent high-level targets such as velocity or waypoint pose. Rotations are encoded with the 6D representation from Zhou et al. (Orsula et al., 27 Sep 2025).

Rewarding is handled through a composite structure,

$r = \sum_i w_i r_i - \sum_j \lambda_j c_j,$

where positive terms reward progress toward the mission objective and penalties discourage unsafe or wasteful behavior. Common penalties include high-frequency action changes, jerk, excessive contact forces, energy or propellant consumption, and environmental hazards. Examples in the paper include soft touchdown and low terminal velocity in landing, balancing scooped regolith against dust generation in excavation, and sparse success bonuses for discrete milestones in screwdriving. Reward functions are intentionally editable in a single task function, enabling studies of reward shaping, curriculum learning, or alternative learning paradigms (Orsula et al., 27 Sep 2025).

The paper reports baselines using PPO, TD3, and DreamerV3. Training uses state-based observations, 512 parallel environments for most tasks, 16 environments for excavation, a single NVIDIA RTX 4090, averages over 3 random seeds, and hyperparameters kept consistent across tasks. The reported main baseline finding is that DreamerV3 is the most sample-efficient and generally achieves the best episodic returns and success rates across nearly all tasks, but training takes about $5.3\times$ longer wall-clock time than PPO under the reported setup (Orsula et al., 27 Sep 2025).

Importantly, several tasks remain unsolved by all three methods, especially solar_panel_assembly and mobile_debris_capture. This is a critical property of a benchmark suite: it establishes not merely solved baselines but open difficulty frontiers. In that sense SRB resembles benchmark designs such as SpaceSense-Bench, where small-component perception and zero-shot generalization remain bottlenecks even for strong baselines, and AstroReason-Bench, where generalist LLM agents underperform specialized solvers across physics-constrained planning problems (Wu et al., 10 Mar 2026, Wang et al., 16 Jan 2026).

6. Empirical findings, sim-to-real transfer, and limitations

The paper’s strongest validation is a successful zero-shot transfer of a learned rover navigation policy to real hardware. In the reported experiment, a Leo Rover was trained in SRB only and then deployed directly in a real-world testbed of lunar-analogue basalt gravel with OptiTrack motion capture for ground-truth localization. Comparing PPO, PPO+LSTM, TD3, and DreamerV3, DreamerV3 achieved the lowest Average Tracking Error in both position and orientation: at 5 cm/s, PPO obtained $13.2$ cm / $7.8^\circ$, PPO+LSTM $11.4$ cm / $g = 0.14 \pm 0.14$0, TD3 $g = 0.14 \pm 0.14$1 cm / $g = 0.14 \pm 0.14$2, and DreamerV3 $g = 0.14 \pm 0.14$3 cm / $g = 0.14 \pm 0.14$4; at 15 cm/s and 25 cm/s, DreamerV3 achieved $g = 0.14 \pm 0.14$5 cm / $g = 0.14 \pm 0.14$6 and $g = 0.14 \pm 0.14$7 cm / $g = 0.14 \pm 0.14$8, respectively (Orsula et al., 27 Sep 2025).

The paper further reports that procedural diversity improves transfer. DreamerV3 trained with one shared procedurally generated terrain is compared with DreamerV3 trained on 512 distinct terrains. At 15 cm/s, the static agent had $g = 0.14 \pm 0.14$9 cm / $g = 1.62 \pm 0.01$0, while the procedural agent had $g = 1.62 \pm 0.01$1 cm / $g = 1.62 \pm 0.01$2, corresponding to 21% lower location ATE and 72% lower orientation ATE (Orsula et al., 27 Sep 2025). This directly supports the framework’s central argument that diversity at scale is an effective route to robustness.

Additional case studies examine environment, object, and morphology diversity. In normalized episodic return, procedural diversity improves locomotion waypoint navigation from 1.00 to 1.03, sample collection from 1.00 to 1.91, and excavation from 1.00 to 9.06. The largest gain arises from varying robot embodiment, suggesting that training across multiple tool geometries materially improves transfer to unseen morphologies (Orsula et al., 27 Sep 2025).

SRB is also used to study adaptive control and end-to-end visuomotor learning. In adaptive control, an adaptive OSC policy that learns motion commands, stiffness, and damping gains improves robustness and safety relative to a rigid IK controller; reported examples include 58% reduction in jerkiness for debris_capture and 56% less jerk with more material excavated in excavation. In visuomotor learning, excavation with two depth cameras yields a 93% increase in scooped regolith compared with a proprioception-only policy (Orsula et al., 27 Sep 2025).

The paper is explicit about limitations. The sim-to-real gap remains real, especially for vision and depth sensing; procedural assets are plausible approximations rather than exact replicas of extraterrestrial sites; and some space-specific physics remain simplified or omitted, including vacuum, extreme thermal cycles, and radiation effects (Orsula et al., 27 Sep 2025). These limitations do not negate the benchmark’s value, but they delimit its role. It is presented as complementary research infrastructure for early-stage development of robust autonomous systems rather than as a replacement for mission-specific high-fidelity simulators (Orsula et al., 27 Sep 2025).

7. Position within the space robotics benchmark landscape

SRB belongs to a broader movement toward standardized evaluation infrastructure in robotics, but its scope is specifically extraterrestrial robot learning. BenchBot contributes a unified observe-and-act interface for benchmarking across photorealistic simulation and real robots; NavBench standardizes reinforcement-learning-based navigation across land, water, and a space-like floating platform; CoBRA formalizes composable descriptions of robots, environments, and tasks for robot selection, base placement, and trajectory optimization; and RobotPerf benchmarks ROS 2 computational graphs in terms of latency, throughput, power consumption, and energy efficiency (Talbot et al., 2020, El-Hariry et al., 20 May 2025, Mayer et al., 2022, Mayoral-Vilches et al., 2023).

Within the more specifically space-oriented literature, several systems provide benchmark-like task platforms or physical analogues without constituting unified learning suites. ReachBot provides a contact-rich mobile manipulation concept for sparse-anchor climbing in microgravity (Schneider et al., 2021). ReCoBot establishes a walking space manipulator with symmetric docking ends for on-orbit servicing scenarios (Scherzinger et al., 2022). Lower-gravity demonstratable testbeds provide compact, repeatable indoor systems for reduced-gravity mobility experiments on steep slopes and uneven terrain (Uno et al., 2023). Hardware-in-the-loop facilities based on UR5e and UR10 manipulators support rendezvous, docking, and debris-capture experiments (Sah et al., 2023). Air-bearing platforms with 7-DOF arms reproduce coupled manipulator-spacecraft dynamics for on-orbit servicing studies (Korczyk et al., 2022). These efforts demonstrate that physical benchmarking in space robotics often centers on locomotion, docking, manipulation, or coupled dynamics in a single regime.

SRB complements these physical and scenario-specific efforts by providing a unified software benchmark suite that spans mobile robotics, fixed-base manipulation, and mobile manipulation under multiple gravity regimes and across many robot embodiments (Orsula et al., 27 Sep 2025). SpaceSense-Bench performs a similar unifying role in spacecraft perception and pose estimation by combining 136 satellite models, RGB, depth, LiDAR, dense 7-class part-level labels, and strict zero-shot evaluation (Wu et al., 10 Mar 2026), while AstroReason-Bench does so for physics-aligned mission planning across communication, revisit, coverage, stereo, and latency optimization problems (Wang et al., 16 Jan 2026). Taken together, these works suggest an emerging benchmark ecosystem for space autonomy in which SRB occupies the role of task- and embodiment-level learning infrastructure.

A plausible implication is that SRB’s lasting significance lies less in any single benchmark score than in its attempt to make space robot learning experimentally comparable, scalable, and extensible. Its central message is that robust extraterrestrial autonomy depends not only on fidelity, but on diversity at scale, and the framework is explicitly engineered to operationalize that principle (Orsula et al., 27 Sep 2025).