Alex Posadas-Nava, Massachusetts Institute of Technology ; August Berne, Massachusetts Institute of Technology ; Giovanni Lavezzi, Massachusetts Institute of Technology ; Kareena Shah, Massachusetts Institute of Technology ; Alejandro Carrasco Aragon, Massachusetts Institute of Technology ; Josiane Uwumukiza, Massachusetts Institute of Technology ; Giacomo Battaglia, Politecnico di Milano; Paolo Panicucci, Politecnico di Milano; Minduli Wijayatunga, University of Illinois at Urbana-Champaign; Victor Rodriguez-Fernandez, Universidad Politécnica de Madrid; Richard Linares, Massachusetts Institute of Technology
https://doi.org/10.64861/QFHT3471
Keywords: 3D reconstruction, 6-DOF Pose Estimation, Generative AI, RPO, Testbed, Software-in-the-loop, SDA
Abstract:
This paper presents the development and initial testing of DreamSat-Bench, a modular rendezvous and proximity operation testbed designed to benchmark AI-based relative navigation techniques. By integrating a software- and hardware-in-the-loop robotic pipeline, the platform enables a seamless transition from software simulation to physical reality. DreamSat-Bench unifies state-of-the-art robotic learning tools such as MuJoCo, Isaac Lab, and LeRobot into a single benchmarking platform and utilizes robotic arms to trace 3D trajectories leveraging contact-rich robotic simulation. The platform allows for extensive customization of orbital environments and lighting to evaluate the simulation-to-reality gap. We demonstrate the testbed’s utility by designing a suite of mission-representative relative motion paths and testing novel techniques for pose estimation based on DreamSat, a generative AI framework for single-view 3D reconstruction. The study also incorporates FoundationPose, a state-of-the-art category-agnostic foundation model, to evaluate zero-shot generalization and high-precision tracking of unseen spacecraft. For the initial testing of DreamSat-Bench, we explore a series of mission-representative simulations that emulate the full chain of a proximity mission, encompassing proximity approach, fixed-point station-keeping, and fly-around characterization. These simulation scenarios are evaluated by comparing the pose estimation accuracy obtained using the ground-truth pose provided by the simulation environment against the pose retrieved by the evaluated AI frameworks and by quantifying the inaccuracies of 3D reconstructions against ground-truth geometric models. DreamSat-Bench aims to provide a rigorous framework for maturing autonomous navigation with unprepared space assets in the absence of prior geometric models.
Date of Conference: September 15-18, 2026
Track: Satellite Characterization