PINGU: Extending Air-Bearing Spacecraft Emulators with Open-Source Actuators and Learned Control for Contact-Rich Proximity Operations

Ricard Marsal I Castan; Akiyoshi Uchida; Aman Arora; Pedro Lima; Matteo El-Hariry; Anrej Orsula; Francesco Grella; Antoine Richard

cs.ROSep 20, 2026

PINGU: Extending Air-Bearing Spacecraft Emulators with Open-Source Actuators and Learned Control for Contact-Rich Proximity Operations

Authors: Ricard Marsal I Castan, Akiyoshi Uchida, Aman Arora, Pedro Lima, Matteo El-Hariry, Anrej Orsula, Francesco Grella, Antoine Richard, +2 more

Organizations: University of Luxembourg · Tohoku University · Georgia Institute of Technology

Abstract

Low-cost planar air-bearing testbeds have matured into a standard proxy for free-flying spacecraft GNC, but they remain largely thruster-only and are rarely equipped for contact-rich, inertia-coupled manipulation. Building on the open-source ATMOS testbed, we contribute a reaction wheel and two force/torque-sensed robotic arms (LEVION) with interchangeable end-effectors, integrated as first-class control actuators through a unified ROS 2 abstraction layer. On top of the software stack we build a reinforcement-learning training environment and digital twin, and a controller that exploits these added degrees of freedom, letting classical optimal controllers and learned policies be swapped on the same hardware without modification. We validate the integrated system, PINGU, across four benchmark tasks: point-to-pose navigation (classical LQR vs. sim-to-real PPO), dynamic disturbance rejection under arm-induced center-of-mass shifts, reaction-wheel momentum stabilization, and force-controlled docking. The results show that these additions extend an ATMOS-class emulator into the contact-rich regime and bridge classical optimal control and reinforcement learning on one reproducible platform.

Figures & tables

Physical properties
	ATMOS [ 5 ]	Astrobee F [ 12 ]	Astrobee G [ 13 ]	Slider [ 3 ]	ESA [ 6 ]	Bob & Charlie [ 10 ]	M-STAR [ 7 ]	JPL SSDT [ 22 ]	ZeroG FP [ 4 ]	PINGU (Ours)
Mass [kg]	16.80	9.58	18.97	4.27	221.67	14.90	–	–	5.32	25
Moment of inertia [kg m 2 ]	0.297	0.162	0.252	0.190	12.223	–	$\mathrm{diag}(1.19,\,1.24,\,1.43)$	–	0.0591	0.72 - 1.35 c
Height [m]	–	0.32	0.32	–	1.025	–	–	–	0.45	0.65
Max. payload [kg]	150	–	–	–	–	–	–	–	20	1360
Actuation

TABLE I: Comparison of PINGU with existing planar/free-flying microgravity testbeds: ATMOS, Astrobee Flight (F), Astrobee Ground (G), Slider, ESA ORGL, Bob & Charlie, M-STAR (6-DoF), JPL SSDT, the UniLu ZeroG floating platform, and our platform. Entries marked – are not reported in the source paper.

Fig. 2: PINGU system overview. Annotated assembly drawing of the complete platform. The chassis is organized into four functional layers from bottom to top: pneumatic base (high-pressure gas bottles and regulator manifold), actuation mid-section (eight cold-gas thrusters, reaction wheel, and Levion mounts), power section (dual LiPo batteries and distribution electronics), and payload deck (exteroceptive sensor board and Levion ).

Fig. 3: New Way 150 mm flat round air bearing (S1015001). Each bearing sustains a 6 $\mathrm{\SIUnitSymbolMicro m}$ aerostatic gap at 5 bar and is rated for 453.5 kg static load. Three bearings are arranged in an equilateral triangle beneath PINGU ’s base plate, providing near-zero friction over the test floor.

Fig. 4: Thruster characterization. Left: Measured force output as thrusters are opened sequentially (1 through 8), showing the progressive pressure drop across the shared supply manifold that reduces individual thrust output under concurrent actuation. Right: CAD model of the thruster box alongside the assembled unit mounted on the PINGU actuation mid-section.

Fig. 5: Reaction wheel. (A) Measured torque-speed envelope of the reaction wheel assembly, showing usable attitude-control authority across the operating speed range. (B) The 20 cm metal-blended PLA disk mounted on the D5312S dual-shaft motor inside the PINGU actuation mid-section.

Fig. 6: Levion design. Planar 2-DOF arm mounted on PINGU ’s actuation mid-section. Each joint is driven by a CubeMars AK80-8 KV60 quasi-direct-drive motor; a Leptrino six-axis F/T sensor at the wrist provides wrench feedback for impedance control. The symmetric bilateral mounting allows the arms to reconfigure the platform’s effective inertia distribution on-the-fly.

Fig. 7: Electronics overview. Two-domain power architecture of PINGU . The control domain (Battery 1, 22.2 V nominal, regulated) supplies the flight controller, compute, and solenoid loads. The actuation domain (Battery 2, direct) supplies the ODrive S1 and CAN-FD arm actuators. Signal paths – including CAN bus, micro-XRCE-DDS bridge, and Ethernet – are shown alongside the power rails. Both domains share a common ground.

Fig. 8: PINGU sensor suite. The four exteroceptive modalities mounted on the dedicated sensor board (Intel RealSense D455, FLIR Firefly, Prophesee event camera, Livox LiDAR), together with representative output from each.

Fig. 9: Two-workspace ROS 2 software architecture of PINGU : the low-level platform workspace exposes all actuators as standard interfaces; the controller-deployment workspace implements classical and learned control logic above that abstraction boundary.

Fig. 10: ROS 2 deployment node graph. External inputs (red) feed the Task Node , which chains four interchangeable modules (blue): a state preprocessor (sensor drivers and velocity estimation), observation formatter (task-specific input assembly), inference runner (policy evaluation), and robot interface (action mapping). The robot interface publishes a unified command to the CMD Mux , which arbitrates between autonomous and manual (joystick) control sources. The Low-Level Controller (green) routes thruster commands to PX4 via an ActuatorMotors uORB message and joint commands to ros2_control (yellow), which manages effort, position, velocity, and trajectory controllers for the reaction wheel and both Levion arms over the CAN bus. Every module is hot-swappable at launch time; the hardware layer requires no modification when the task, sensor, or policy changes.

Fig. 11: PINGU digital twin in the Space Robotics Bench. Thousands of parallelized copies of the laboratory environment running simultaneously on a single GPU inside IsaacLab. Each instance contains a full PINGU model (air-bearing platform with two Levion arms) and two UR10 manipulator mounted to the ceiling and wall rails, replicating the physical laboratory layout. Massively parallel rollouts with domain randomization enable the sim-to-real transfer results reported in Section IV .

Fig. 12: Levion arm configurations. Three static arm poses used throughout the experiments to vary the platform’s effective moment of inertia: Rest (arms extended outward at $90^{\circ}$ , maximizing inertia); Closed (arms tucked toward the body, minimizing inertia); and Side (both arms extended to one side, producing an asymmetric mass distribution). The configurations are held fixed during each run and are not part of the controller’s action space; they serve as a structured way to stress-test controller robustness across different inertial regimes.

Fig. 13: Point-to-pose navigation. Left: hardware trajectories of PINGU navigating from four random initial poses (colored dots) to a fixed target pose and heading (arrow), with arrowheads marking direction of travel. Right: corresponding distance error (top) and heading error (bottom) over time for the same four runs, both converging to near zero well within the 60 s trial.

Methods	Point-to-Pose Navigation
Methods	$e_{p}[m]\downarrow$	$e_{o}[rad]\downarrow$
(a) All actuators
Sim	0.0022 ( $\pm$ 0.0015$$ )	0.0044 ( $\pm$ 0.0033$$ )
Sim + DR	0.0124 ( $\pm$ 0.0191$$ )	0.0129 ( $\pm$ 0.0124$$ )
PPO (Real)	0.0081 ( $\pm$ 0.0006$$ )	0.0257 ( $\pm$ 0.0146$$ )
(b) Thrusters and reaction wheel

TABLE II: Point-to-pose navigation. Final position error

e_{p}

and heading error

e_{o}

comparing the learned PPO controller (Sim, Sim+DR, and Real) against the classical LQR controller (Real only), across three actuator subsets and three static Levion arm poses (Side, Rest, Closed). The all-actuators configuration has no LQR baseline as the arms are part of the PPO action space.

Fig. 14: Dynamic disturbance rejection. Top: heading error $\theta$ over time for four representative hardware runs, converging toward zero despite the continuous CoM disturbance. Middle: corresponding distance error to the target pose, converging with bounded residual oscillations driven by the disturbance. Bottom: commanded shoulder/elbow joint positions for the same four runs, showing the disturbance generator’s continuous $\pm 0.8$ rad oscillation throughout each 60 s trial. Frame sequence: a representative run, with the platform navigating to the target pose while its arms swing autonomously.

Methods	Dynamic Disturbance Rejection
Methods	$e_{p}[m]\downarrow$	$e_{o}[rad]\downarrow$
Sim	0.0070 ( $\pm$ 0.0054$$ )	0.0074 ( $\pm$ 0.0108$$ )
Sim + DR	0.0156 ( $\pm$ 0.0161$$ )	0.0340 ( $\pm$ 0.0574$$ )
Real	0.0889 ( $\pm$ 0.0405$$ )	0.0346 ( $\pm$ 0.0179$$ )

TABLE III: Dynamic disturbance rejection. Final position error

e_{p}

and heading error

e_{o}

for the PPO policy navigating to a fixed target pose under continuous, unobserved CoM disturbance from arm motion, evaluated in simulation, in simulation with domain randomization (DR), and on hardware.

Fig. 15: Momentum damping in the Rest arm configuration. Top: magnitude of the platform’s angular velocity over time across four representative hardware stabilization runs; dotted lines mark half the initial angular velocity and diamond markers indicate each run’s half-time

t_{1/2}

. Bottom: corresponding commanded reaction-wheel torque for the same four runs, showing the PPO policy applying progressively less braking torque as the angular velocity decays toward zero.

Methods	Stabilization
Methods	$\phi[rad/s]\downarrow$	$t_{\text{half}}[s]\downarrow$
(b) Rest
Sim	0.0467 ( $\pm$ 0.0153$$ )	3.7975 ( $\pm$ 1.6068$$ )
Real	0.0228 ( $\pm$ 0.0229$$ )	6.377 ( $\pm$ 4.941$$ )
(c) Closed
Sim	0.0432 ( $\pm$ 0.0150$$ )	3.8773 ( $\pm$ 1.6457$$ )

TABLE IV: Momentum damping / stabilization. Residual angular velocity

\phi

and half-time

t_{1/2}

of the reaction-wheel-only PPO stabilization policy, evaluated in simulation and on hardware across three Levion arm configurations (Side, Rest, Closed) spanning the platform’s range of effective moment of inertia.

Fig. 16: Force-controlled docking. Top: frame sequence of a representative run, with the platform’s trajectory approaching the wall-mounted ArUco target. Middle: magnitude of the F/T-sensed contact force at the left (solid) and right (dashed) wrist over four repeated runs, showing intermittent contact spikes once the arms reach the wall. Bottom: magnitude of the platform’s linear velocity for the same four runs, peaking mid-approach and decaying to near zero as the arms’ compliant contact dissipates the platform’s residual kinetic energy.

Hyperparameter	PPO (MLP)	PPO (GRU)
Number of envs	4096
Number of steps per iteration	16
Number of learning epochs	5
Number of mini-batches	4
Clip range	0.2
Entropy coefficient	0.005

TABLE V: PPO hyperparameters for the Dynamic Disturbance Rejection policies, for the feed-forward (MLP) and recurrent (GRU) variants. Shared values span both columns.

(a) Point-to-Pose Navigation — all actuators
Method	$e_{p}\,[\text{m}]\downarrow$	$e_{o}\,[\text{rad}]\downarrow$
PPO (Sim)	0.0022 ( $\pm$ 0.0015$$ )	0.0044 ( $\pm$ 0.0033$$ )
PPO + DR (Sim)	0.0274 ( $\pm$ 0.0760$$ )	0.0192 ( $\pm$ 0.0187$$ )
PPO + DR (Real)	0.0584 ( $\pm$ 0.0412$$ )	0.0560 ( $\pm$ 0.0151$$ )
PPO-GRU + DR, 1000 ep (Sim)	0.0209 ( $\pm$ 0.0344$$ )	0.0277 ( $\pm$ 0.0255$$ )
PPO-GRU + DR, 1000 ep (Real, chkpt. 1)	0.0862 ( $\pm$ 0.0862$$ )	0.0443 ( $\pm$ 0.0307$$ )

TABLE VI: All-actuator navigation and dynamic disturbance rejection — full ablation. Final position error

e_{p}

and heading error

e_{o}

for the two studies in which the arms are actuated (all-actuators navigation) or continuously animated as a disturbance (dynamic rejection), so no static arm pose applies. For the all-actuators policy we also sweep the recurrent training budget (1000 vs. 2000 epochs); for the hardware entries we report multiple training seeds / deployment runs. Lower is better.

Term	Expression	Weight	Scalar	Description
(a) Point-to-Pose Navigation	$r_{t}^{\text{point-pose-nav}}=r_{t}^{\text{pose}}+r_{t}^{v,\omega}+r_{t}^{\text{bnd}}$
Pose error	$\alpha_{\text{pose}}(\text{exp}(-d_{p}/\lambda_{p})\times\text{exp}(-d_{h}/\lambda_{h}))$	$\alpha_{\text{pose}}=1.0$	$\lambda_{p},\lambda_{h}=1.0$	Precise spatial convergence and alignment to target.
Linear Velocity	$\alpha_{\text{v}}\text{clip}(v-v_{\text{min}},0,v_{\text{max}}-v_{\text{min}})$	$\alpha_{\text{v}}=-0.05$	$v_{\text{min}}=0.5,v_{\text{max}}=2.0$	Velocity regulation and movement encouragement.
Angular Velocity	$\alpha_{\omega}\text{clip}(\omega-\omega_{\text{min}},0,\omega_{\text{max}}-\omega_{\text{min}})$	$\alpha_{\omega}=-0.05$	$\omega_{\text{min}}=0.5,\omega_{\text{max}}=20.0$	Control of rotational stability.
Boundary	$\alpha_{\text{bnd}}\text{exp}(-d_{b}/\lambda_{b})$	$\alpha_{\text{bnd}}=-10.0$	$\lambda_{b}=1.0$	Safety constraint to stay within workspace.
(b) Reaction-Wheel Stabilization (Momentum Damping)	$r_{t}^{\text{stab}}=r_{t}^{\omega}+r_{t}^{\text{sat}}+r_{t}^{\text{use}}$

TABLE VII: Reward terms used to train the learned control policies. Each block lists a task-specific reward, given as the weighted sum of its rows. Block (c) collects the platform-regularization penalties added to the Point-to-Pose Navigation policy (a); as noted per row, individual terms apply only to configurations that include the corresponding actuators (reaction wheel, arms). The reaction-wheel Stabilization policy (b) actuates only the wheel and is therefore regularized by just the reaction-wheel saturation and usage terms, which share the definitions and weights of the matching rows in (c). Distances

d_{\bullet}

denote the absolute error between the current and target quantity; negative weights indicate penalties.

(a) Thrusters and reaction wheel
	Side		Rest		Closed
Method	$e_{p}\,[\text{m}]\downarrow$	$e_{o}\,[\text{rad}]\downarrow$	$e_{p}\,[\text{m}]\downarrow$	$e_{o}\,[\text{rad}]\downarrow$	$e_{p}\,[\text{m}]\downarrow$	$e_{o}\,[\text{rad}]\downarrow$
PPO (Sim)	0.0021 ( $\pm$ 0.0010$$ )	0.0038 ( $\pm$ 0.0043$$ )	0.0031 ( $\pm$ 0.0026$$ )	0.0020 ( $\pm$ 0.0022$$ )	0.0034 ( $\pm$ 0.0023$$ )	0.0018 ( $\pm$ 0.0013$$ )
PPO + DR (Sim)	0.0155 ( $\pm$ 0.0423$$ )	0.0193 ( $\pm$ 0.0175$$ )	0.0204 ( $\pm$ 0.0624$$ )	0.0406 ( $\pm$ 0.0513$$ )	0.0154 ( $\pm$ 0.0360$$ )	0.0122 ( $\pm$ 0.0122$$ )
PPO + DR (Real)	0.0115 ( $\pm$ 0.0076$$ )	0.0302 ( $\pm$ 0.0088$$ )	0.0246 ( $\pm$ 0.0155$$ )	0.0313 ( $\pm$ 0.0139$$ )	0.0406 ( $\pm$ 0.0227$$ )	0.0121 ( $\pm$ 0.0066$$ )
PPO-GRU (Sim)	0.0137 ( $\pm$ 0.0112$$ )	0.0047 ( $\pm$ 0.0039$$ )	0.0220 ( $\pm$ 0.0217$$ )	0.0047 ( $\pm$ 0.0036$$ )	0.0045 ( $\pm$ 0.0055$$ )	0.0040 ( $\pm$ 0.0033$$ )

TABLE VIII: Point-to-Pose Navigation - static-arm ablation. Final position error

e_{p}

and heading error

e_{o}

for the reduced-actuator subsets, evaluated across the three static Levion arm poses ( Side , Rest , Closed ) of Fig. 12 . Each subset is ablated over policy architecture (feed-forward PPO vs. recurrent PPO-GRU), domain randomization (DR), and deployment (Sim vs. Real hardware). Entries are mean

\pm

std over evaluation episodes; lower is better.

Explore similar work

Sep 29, 2026cs.RO

Wrench-ACT: Enhancing Robot Policies for Contact Rich Behavior Using Direct Wrench Control

While contact-rich manipulation requires deliberate regulation of interaction forces, recent approaches to robot manipulation learning predominantly represent actions as target positions or poses. Even methods that incorporate force sensing either use it solely as an observation or, when predicting forces as part of the output, rely on a hybrid force controller. In this paper, we propose an imitation learning policy that predicts wrenches as its sole action output for direct use by a pure force controller. Our studies suggest that force-domain imitation learning depends critically on data collection, with force-feedback teleoperation improving policy performance by capturing the operator's deliberate force regulation. Using Action Chunking with Transformers (ACT) as the base architecture, we train single-task models on bilateral wrench demonstrations and evaluate them on five contact-rich manipulation tasks. The wrench policy matches or outperforms position-based baselines across all tasks, with gains varying according to the degree of deliberate force regulation each task requires. Cross-condition ablations show that the bilateral data collection interface and the wrench action space each contribute independently to performance. To support further research, we will release over 1000 wrench-action demonstrations spanning these tasks on a companion website upon publication.

Johannes Hechtl, Yannik Blei, Simon Ball +5

Siemens Research and Predevelopment · Department of Computer Science and Artificial Intelligence, University of Technology Nuremberg, Germany

Robotic Manipulation Policies Contact Wrench Estimation Contact-Rich Manipulation Robot Policies Scalable Robot Learning Temporally Coherent Imitation Learning+3

Ask PDF

May 23, 2026cs.RO

Towards Low-Gravity Planetary Exploration using Reinforcement Learning for Walking, Jumping, and In-flight Attitude Control

This paper presents reinforcement learning (RL) policies for dynamic quadrupedal locomotion in planetary exploration scenarios. Building on a taskoptimized quadruped with a 5-bar leg design, we develop RL policies for walking, vertical jumping, forward jumping, and in-flight attitude control, explicitly tailored to the reduced gravity on Mars. These policies jointly enable such robots to overcome obstacles larger than themselves through coordinated jumping and precise in-flight reorientation for safe landings. We demonstrate Sim2Real transfer of the attitude control policy on the Olympus quadruped through single-axis reorientation tests, while all locomotion policies are validated in simulation. A complete Mars exploration mission scenario demonstrates coordinated policy deployment across challenging terrain. Experimental results show 90° attitude reorientation in 2.6 seconds, with simulations demonstrating 3.1 meter vertical jumps and 3.9 meter forward jumps under Martian gravity conditions. - Supplementary video: https://www.youtube.com/watch?v=qlSJ3P87A4A

Jørgen Anker Olsen, Kostas Alexis

The authors are with the Autonomous Robots Lab, NTNU, O.S. Bragstads Plass 2D, 7034, Trondheim, Norway.

Planetary Exploration Reinforcement Learning Control Active Flow Control Exploration

Ask PDF

Jun 9, 2026cs.RO

IMPACT: Learning Internal-Model Predictive Control for Forceful Robotic Manipulation

Real-world robotic manipulation tasks often involve forceful interactions with the environment, such as using tools of varying weights, transporting objects with different masses, and performing contact-rich tasks like table wiping. Previous learning-based approaches typically employ imitation learning policies that output target end-effector poses tracked by low-level impedance controllers. In these systems, forceful interactions are either implicitly realized through steady-state tracking errors or explicitly commanded using wrist force/torque or tactile sensors. However, implicit approaches generalize poorly across object weights, while explicit approaches require specialized hardware and increase system complexity. In this work, we propose IMPACT, a framework that decouples these forceful tasks into task-planning and internal-model-based predictive control. Extensive simulation and real-world experiments demonstrate that the proposed framework achieves higher success rates and improved generalization to unseen object weights, as well as better safety and energy efficiency.

Jiawei Gao, Chaoqi Liu, Peilin Wu +2

Harvard University · Stanford University

Robotic Manipulation Impedance Control End-Effector Pose Model Predictive Control Implicit Impact Analysis

Ask PDF