Space logistics and transportation.
Reaching a destination in space is a velocity budget, and the whole chain, from launch through orbital transfer, rendezvous, capture, servicing, and return, becomes tractable when reusable infrastructure and autonomous in-space transport pay it down. This is the instrument suite: each transport mode running on its real equations, and the autonomy that unites them, so the capabilities can be tested, applied, and deployed toward an open, decentralized space economy.
Technical Report TR-2026-17 · Institute for Physical AI @ BMI
The autonomy that unites the chain.
One shape repeats across every link: an estimator resolving state, and a controller holding a proven corridor. The clearest case is the hardest, a rendezvous with a body that neither holds still nor helps.
A Kalman filter resolves the tumbling target's pose from noisy returns (the blue covariance shrinks as it converges); the chaser holds the certified corridor, matches the spin, and latches inside the capture envelope. Switch to MPPI to watch a real optimal controller sample hundreds of trajectories and choose its path, or to Learned to fly the approach with a 1,282-parameter policy distilled from that planner and run on-device (released on Hugging Face), or to human-in-loop, where the control lag opens the miss. Illustrative dynamics on a real filter, a real planner, and a real learned policy.
Seven instruments, real physics.
Each runs on its governing equations, in the browser, on the device. Open any one, tune it, and read where the gate is autonomy and control rather than materials.
Compose the journey
Pick a destination and stack reusable infrastructure and autonomous in-space transport against the velocity budget. The mass-ratio math is the exact rocket equation; a segment a single vehicle could never fly is marked so, and the payload it delivers is read as a real mass fraction.
Open the instrument ↗On-orbit capture
A chaser estimates a tumbling, non-cooperative target's pose from noisy returns with a Kalman filter, holds a certified approach corridor, matches the spin, and latches. Toggle autonomous against human-in-loop to see why the loop runs on the vehicle.
Open the instrument ↗Monocular pose
A tumbling, non-cooperative target gives no telemetry. From one camera's noisy keypoints the full six-degree pose is recovered by solving perspective-n-point live, and scored against the truth. This is the SPEED benchmark task, running on the device: rotation error to a degree or two, translation to a few centimetres.
Open the instrument ↗Servicing & assembly
Past capture, the chain continues: a manipulator inserts a replacement unit into a berthed satellite. Command an exact position and it jams on the rim; command a gentle force and let the chamfer seat it. Compliance, not precision, closes the last millimeter.
Open the instrument ↗Skyhook · rotovator
The momentum-exchange tether on real orbital mechanics: catch and release velocities from vis-viva, the payload's true transfer orbit after release, and the tether-to-payload mass from the constant-stress taper equation. Change the material and watch it become buildable.
Open the instrument ↗Maglev sled
The electromagnetic ground assist: track length grows with the square of exit speed and is capped by the g-load the payload survives. Dial the exit speed and g-limit and read the track length, the peak power, and the coil-switching rate.
Open the instrument ↗Launch loop
An elevated track held up by a fast internal rotor: an unstable equilibrium stabilized span by span. Turn stabilization off, or push the rotor past its control margin, and it buckles. The frame is ordinary steel; the stiffness is the controller.
Open the instrument ↗One stack, every link.
Read across the suite and the same problem recurs: a launch handoff at speed, a momentum-exchange catch, a non-cooperative capture. Each is an estimator and a certified controller running on the vehicle. One autonomy stack, priced in Δv and joules, serves every link in the logistics chain. That is the research this track develops, and the capability the space economy needs to open.
◆ RPO-Bench · the open leaderboard ↗ · a deterministic, reproducible benchmark for capture and pose; run it headless or live in your browser and try to beat the baselines.
orbital-logistics on GitHub ↗· capture policy on Hugging Face ↗ · the instrument suite, a headless physics core (with an MPPI planner), and a distilled on-device capture policy.
↓ Whitepaper · PDFRead online◆ Living paper
Take the course · PAI-260 →← Back to the labBuild this with us →