Under Review (2026)

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

Ricard Marsal I Castan1, Akiyoshi Uchida2, Aman Arora1, Pedro Lima1, Matteo El-Hariry1, Anrej Orsula1, Francesco Grella1, Antoine Richard1, Cedric Pradalier3, Miguel A. Olivarez-Mendez1
1University of Luxembourg 2Tohoku University 3Georgia Institute of Technology
Paper ArXiv Code Task Demos 🤗 Models
PINGU Platform Overview

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.

Key Capabilities

Heterogeneous Actuation

Co-integrates an 8-thruster planar RCS, a high-torque brushless reaction wheel, and dual 2-DOF LevionArms with 6-axis force/torque wrist sensors on a single chassis.

Unified ROS 2 Stack

Exposes all hardware through a hardware-abstraction command multiplexer, allowing model-based (PID, LQR, MPC) and learned (PPO) policies to swap seamlessly.

Sim-to-Real Twin

Supported by a parallelized GPU simulation inside Isaac Lab (Space Robotics Bench) with extensive domain randomization over mass, inertia, and external forces.

System & Mechanical Design

PINGU features a vertically-layered, modular chassis (40cm x 40cm x 65cm) constructed from aluminum T-slot extrusions and polycarbonate panels. The system is split into four distinct functional layers designed to isolate structural, pneumatic, electronic, and robotic subsystems:

Mechanical Stack Assembly Diagram

Figure 1. Mechanical Stack: (1) Pneumatic base with porous-graphite air bearings; (2) Actuation mid-section with reaction wheel and motors; (3) Power section; (4) Payload deck with sensors and robotic arms.

Four Stacked Layers

  • Pneumatic Base: Carries three 150mm flat round porous-graphite air bearings arranged in an equilateral triangle. Using a 5-bar regulated supply, they sustain a 6μm aerostatic gap, supporting up to 1360 kg total load.
  • Actuation Mid-Section: Houses the 8 cold-gas thruster solenoid valves, the 20cm diameter reaction wheel (metal-blended PLA disk with AMT212B encoder), and dual CubeMars AK80-8 motors.
  • Power Section: Integrates dual isolated LiPo domains to isolate sensitive control hardware from high-current motor transients.
  • Payload Deck: Holds the Jetson computing stack and exteroceptive sensor board.

Dual-Domain Electronics & Sensing

A Holybro Pixhawk Jetson Baseboard integrates a Pixhawk 6C flight controller and an NVIDIA Jetson Orin NX.

  • Control Domain (Battery 1, 24V Regulated): Powers computing, exteroceptive sensors, and thruster solenoids.
  • Actuation Domain (Battery 2, 22.2V Direct): Powers ODrive S1 motor controllers and CubeMars CAN-FD actuators.
  • Exteroceptive Sensor Suite: Includes an Intel RealSense D455 RGB-D camera, FLIR Firefly camera, Prophesee Event camera, and Livox Mid-range LiDAR, linked via onboard Gigabit Ethernet.
PINGU Electronics Block Diagram

Figure 2. Electronics Architecture: Split power domain routing with CAN bus communication lines and Ethernet signal paths.

Actuator Characterization

To build a high-fidelity digital twin and enable accurate control, we characterized both the thrusters and reaction wheel assemblies.

Thruster characterization plots

Thruster Manifold Coupling

Thruster box and pressure drop: Solenoids fire up to 500Hz (modulated at 10Hz in software). Firing multiple valves concurrently causes a drop in individual thrust output due to sharing the same supply manifold, which must be randomizing-modeled.

Reaction wheel characterization plots

Reaction Wheel Envelope

Reaction wheel torque envelope: The 20cm metal-blended PLA disk delivers continuous, non-impulsive attitude torque. The measured torque-speed envelope characterizes the operational control authority limit.

Software Stack & Simulation

Two-Workspace ROS 2 Layer

PINGU runs a dual ROS 2 workspace architecture to strictly isolate drivers from controllers:

Low-Level Workspace: Runs hardware-facing nodes. Solenoid thruster inputs map directly to PX4 offboard actuator motors, and robotic arm/reaction wheel joints communicate with `ros2_control` over the shared CAN-FD bus.

Controller-Deployment Workspace: Runs a high-level RoboRAN task node. It preprocessing sensor telemetry, runs a 10Hz inference loop on PPO policies (MLP/GRU) via ONNX, or routes PID/LQR commands, feeding a safe command multiplexer.

ROS 2 Node Graph diagram

Figure 3. Node Graph: The state processor, observation formatter, inference runner, and low-level controllers linked via ROS 2.

Space Robotics Bench digital twin in Isaac Lab

Figure 4. Parallelized Digital Twin: Space Robotics Bench (SRB) environment running thousands of instances simultaneously inside Isaac Lab.

Space Robotics Bench (Digital Twin)

The digital twin is implemented within the Space Robotics Bench (SRB) on top of NVIDIA Isaac Lab and Isaac Sim.

It models multi-body dynamics, binary thruster characteristics, reaction wheel speed limits, and the active 2-DOF LevionArms. Massively parallel rollouts on a single GPU enable quick reinforcement learning training. To close the sim-to-real gap, domain randomization is applied to base mass (±5kg), horizontal Center-of-Mass offset (0.1m), and constant external bias wrench (representing table tilt and thruster imbalance).

Experimental Tasks & Videos

Select a task below to view details, experimental results, and real hardware videos.

Point to pose navigation trajectory

Point-to-Pose Navigation: Trajectories from 4 random starting states converging to target (left), and error plots (right).

Point-to-Pose Navigation

The foundational maneuver requires navigating to a target coordinate and heading using three actuator subsets: thrusters only, thrusters + reaction wheel, and all actuators (with active arms). Static arm poses (Side, Rest, Closed) test varying inertial conditions.

Configuration Method Position Error [m] ↓ Heading Error [rad] ↓
All Actuators PPO (Real) 0.0081±0.0006 0.0257±0.0146
Thruster + RW (Side) PPO (Real) 0.0547±0.0422 0.0275±0.0154
LQR (Real) 0.0970±0.0574 0.0306±0.0111
Thruster Only (Side) PPO (Real) 0.0241±0.0160 0.0335±0.0155
LQR (Real) 0.0830±0.0248 0.0229±0.0234

Hardware Video: PINGU correcting translation/heading while arms oscillate autonomously.

Dynamic Disturbance Rejection

The two robotic arms oscillate autonomously at ±0.8 rad, displacing the Center of Mass (CoM). Because arm motion is unobserved by the controller, it represents an exogenous dynamic perturbation that must be rejected implicitly from base feedback.

Method Deployment Position Error [m] ↓ Heading Error [rad] ↓
PPO-GRU Simulation (Sim) 0.0070±0.0054 0.0074±0.0108
PPO-GRU + DR Sim + Rand (Sim) 0.0156±0.0161 0.0340±0.0574
PPO-GRU + DR Real Hardware 0.0889±0.0889 0.0346±0.0179
Dynamic Disturbance Rejection plot

Figure 5. Telemetry: Error profiles and commanded joint position oscillations over a 60s trial.

Base Stabilization
Reaction Wheel Spin

Reaction Wheel Stabilization: Left shows base stabilization; Right shows reaction wheel assembly operating.

Momentum Dumping & Stabilization

PINGU is spun with a random initial angular rate between 0.1 to 0.5 rad/s. The task is to stabilize the base's attitude using only the reaction wheel. This is evaluated under different arm poses to test wheel response under varying base moments of inertia.

Arm Configuration Deployment Residual Yaw Rate [rad/s] ↓ Half-time t1/2 [s] ↓
Side (Asymmetric) Simulation 0.0429±0.0148 4.0773±4.7499
Real Hardware 0.0594±0.0304 4.8450±1.8730
Rest (Max Inertia) Simulation 0.0467±0.0153 3.7975±1.6068
Real Hardware 0.0228±0.0229 6.3770±4.9410
Stabilization plots

Figure 6. Telemetry: Angular velocity decay profile (top) and commanded wheel brake torque (bottom) across four runs.

Force-Controlled Docking

This integrates PINGU's full stack:

1. The onboard RealSense D455 camera detects a wall-mounted ArUco marker to estimate the target pose relative to the robot.
2. The PPO point-to-pose controller drives the platform towards the docking wall.
3. Leptrino six-axis F/T sensors on the wrists register the contact onset.
4. Base propulsion stops, letting momentum carry the robot. The contact is safely damped by joint-space virtual spring-damper impedance loops in the LevionArms, absorbing kinetic energy without base rebound.

Docking Approach
ArUco Visual Tracker
Wall Docking Contact
Docking forces plots

Figure 7. Contact Forces: Telemetry showing F/T contact force spikes during compliance damping and linear velocity decay.

BibTeX

@misc{castan2026pinguextendingairbearingspacecraft,
      title={PINGU: Extending Air-Bearing Spacecraft Emulators with Open-Source Actuators and Learned Control for Contact-Rich Proximity Operations}, 
      author={Ricard Marsal I Castan and Akiyoshi Uchida and Aman Arora and Pedro Lima and Matteo El-Hariry and Anrej Orsula and Francesco Grella and Antoine Richard and Cedric Pradalier and Miguel A. Olivarez-Mendez},
      year={2026},
      eprint={2609.23554},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2609.23554}, 
}