SmallSatSim: A GPU-Accelerated Microgravity Robotics Toolkit for Planning, Control, and Policy Learning
Authors: David Schwartz, Alexander Hansson, Sabrina Bodmer, David Sternberg, Oliver Jia-Richards, Keenan Albee
Organizations: ETH Zürich, Rämistrasse 101, 8092 Zürich, Switzerland · Jet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove Drive, Pasadena, CA 91011, USA · University of Michigan, 500 South State Street, Ann Arbor, MI 48109, USA · University of Southern California, 3551 Trousdale Parkway, Los Angeles, CA 90089, USA
Microgravity rendezvous and close proximity operations (RPO) is a growing area of interest for applications spanning in-space assembly and manufacturing (ISAM), orbital debris remediation, and small body exploration. Developing autonomy for these operations requires integrating dynamics simulation with task definition, control and learning algorithms, robustness testing, and evaluation. We present SmallSatSim, an open-source toolkit for developing autonomy algorithms for robots operating in microgravity environments. Built on MuJoCo, SmallSatSim provides a common experiment and task abstraction for spacecraft models, planners, controllers and policies, actuator effects, disturbances, and evaluation, allowing model-based and learning-based approaches to operate on the same problem definitions. The framework combines conventional MuJoCo execution for model-based control with vectorized JAX/MJX execution for massively parallel policy learning. We demonstrate SmallSatSim through Monte Carlo experiments on model-based control under actuator perturbations, five-seed PPO and SAC training under nominal and randomized dynamics with out-of-distribution evaluation, contact-rich rendezvous and docking, and GPU scaling experiments.
Figures & tables
Systems property
SmallSatSim MuJoCo/MJX
Isaac Lab Newton
Isaac Lab PhysX
Massively parallel GPU simulation
•
•
•
JAX/XLA JIT-compilation
•
–
–
Automatic differentiation
•
∘
–
TABLE I: Comparison of selected simulation stacks for robot learning. • denotes native/full support, ∘ partial or solver-dependent support, and – no native support.
Capability
SmallSatSim
SRB
mjorbit
Basilisk / BSK-RL
Astrobee Gazebo/FSW
Physical scope
Local microgravity simulation
•
•
•
•
•
Orbit/environment simulation
–
–
•
•
–
Sensor simulation
–
•
•
•
•
Simulation systems
Accelerator-parallel simulation
•
•
•
–
–
TABLE II: Comparison of simulation and autonomy-system capabilities across representative space-robotics frameworks. • denotes a reusable framework capability, ∘ denotes a demonstrated example that is not provided as a reusable framework component, and – denotes a capability not provided or evidenced in the public framework. Per-environment randomization means that each parallel environment can use independently randomized spacecraft and actuator parameters.
Fig. 2: SmallSatSim ’s modular experiment architecture. A common task interface connects spacecraft models, planners, controllers or policies, failures, disturbances and randomization, and CPU or GPU simulation backends.
Fig. 3: Policy-learning and robustness-evaluation workflow used in our experiments. Policies are trained under nominal or randomized dynamics and evaluated on common seeded nominal, ID, OOD, and stress scenarios.
Fig. 4: Lunar Gateway inspection by Astrobee, using the NominalMPCController and MissionPlanner inspection planner.
Fig. 5: Monte Carlo robustness distributions for model-based control under actuator perturbations and disturbances.
Condition
Success
Final pos. error [m]
Final att. error [rad]
Control effort [N s]
Nominal
100/100
0.074 [0.037, 0.174]
0.229 [0.228, 0.231]
10.97 [8.48, 12.82]
Disturbance
100/100
0.068 [0.033, 0.151]
0.229 [0.227, 0.231]
10.85 [8.28, 12.92]
Faulty valve
100/100
0.076 [0.038, 0.176]
0.228 [0.222, 0.230]
11.19 [8.70, 13.05]
Saturated thrust
100/100
0.119 [0.047, 0.185]
0.229 [0.227, 0.231]
12.62 [9.28, 14.29]
Stuck-off
100/100
0.128 [0.086, 0.164]
0.230 [0.228, 0.237]
14.00 [10.91, 19.23]
Stuck-on
45/100
0.185 [0.118, 0.191]
0.263 [0.237, 0.410]
33.15 [14.37, 45.70]
TABLE III: Model-based robustness over 100 trials per condition. Errors and control effort are reported as median [Q1, Q3], where Q1 and Q3 denote the 25th and 75th percentiles.
Fig. 6: Policy learning across five training seeds and evaluation under nominal, randomized-initial-state, thrust-variation, disturbance, and combined conditions. Shaded bands show one standard deviation across training seeds; error bars show 95% Student- t confidence intervals across seeds.
Fig. 7: Rendezvous and docking with compliant contact at the Lunar Gateway.
Fig. 8: Aggregate rollout throughput with increasing numbers of parallel environments on an NVIDIA A100-SXM4-40GB GPU.
# Envs
Env Steps/s
Aggregate Sim s/s
Per-env Sim s/s
Speedup
1
4.50×103
225.1
225.1
1.0×
8
3.11×104
1556.0
194.5
6.9×
32
1.21×105
6041.8
188.8
26.8×
128
4.73×105
23635.1
184.6
105.0×
512
1.69×106
84424.3
164.9
375.1×
1024
3.30×106
164991.6
161.1
733.0×
TABLE IV: GPU rollout performance using the free-flyer dynamics backend on an NVIDIA A100-SXM4-40GB GPU. Values are means over 10 repeats.
This paper presents a general framework for simulating multi-body space robots with contact. We bring efficient, large-scale robot simulation to in-space servicing, assembly, and manufacturing applications. First, we perform an empirical trade study of methods for coupling orbit propagation with existing robotics simulation frameworks. Next, we present mjorbit, a general, flexible, and performant framework built on the MuJoCo engine widely used in robotics, to which we add key spacecraft dynamics, actuators, and sensors. We provide a low-latency C++ CPU backend and a high-throughput GPU backend behind a simple Python API. We demonstrate mjorbit by solving several realistic on-orbit case studies with both model-predictive control and reinforcement learning. Open-source code and examples are available at: https://johnzhang3.github.io/mjorbit/
John Z. Zhang, Joris Verhagen, Fausto Vega +2
Department of Aeronautics and Astronautics, Massachusetts Institute of Technology · Division of Robotics, Perception and Learning, KTH Royal Institute of Technology
Developing learning-based autonomy for microrobotic manipulation is difficult because physical trials are slow, variable, and poorly suited to large-scale training. We present MicroPush, a lightweight simulation and benchmarking environment for pushing and assembly with magnetic rolling microrobots. The simulator combines overdamped planar dynamics with calibrated rolling actuation, reduced-order contact handling, short-range interaction corrections, and optional background flow, while providing common interfaces for planning, model-based control, reinforcement learning (RL), interactive visualization, and headless evaluation. In no-render tests, MicroPush remains substantially faster than real time, reaching 37.0x real time with 100 passive particles. We further evaluate the simulator through zero-shot sim-to-real policy transfer. A Residual RL-MPC policy is trained entirely in MicroPush, frozen after simulation training, and deployed on the physical system without real-world policy fine-tuning. A 20 μm magnetic microrobot pushes a similarly sized silica particle along a figure-eight trajectory with 1.91 μm median and 4.65 μm 95th-percentile path error, and the same autonomy stack performs sequential line, triangle, and six-particle hexagonal assembly. These results show that a deliberately reduced-order simulator can provide an efficient training and development substrate for sim-to-real microrobotic autonomy.
Yanda Yang, Zion Lee, Sambeeta Das
Department of Mechanical Engineering, University of Delaware, Newark, DE 19716, USA
This work demonstrates the use of a fast simulation-based training framework to learn control policies using reinforcement learning. The hardware requirements are minimal and the simulation-learned policies transfer reliably to the real world. With this setup, we show that a quadruped robot can learn to traverse diverse real-world terrains reliably---including in a lunar-analog facility---after less than 20 minutes of training on a single consumer-grade GPU. These results provide a glimpse of a future where reinforcement learning is a crucial tool in developing more capable and robust space robotic systems. Video demonstrations are available at https://youtu.be/vUFJiZdJaQ0.
Abhishek Naik, Michael Wu, Michael O'Sullivan +2
National Research Council Canada · University of Waterloo · University of Ottawa +1