Taking Learn to Teach to Sea
On April 2, 2026 we took Digit aboard the U.S. Navy’s Stiletto and ran our Learn to Teach (L2T) locomotion policy inside the ship’s mission bay. It was the first time the policy left the lab and campus terrain for a real vessel.
Digit walking toward the camera in the Stiletto’s mission bay, running the L2T policy.
Why a ship
L2T trains a privileged teacher and a proprioceptive student together in simulation and transfers the student to hardware zero-shot. On land we had already covered rocky, grassy, sandy, and slippery ground, plus pushes and payloads. A ship is a different kind of test: the floor is steel with drainage plates and tie-down tracks, the space is tight and full of gear, and once the ship is underway the deck itself moves under the robot. None of that appears in the training terrains, so the question was simply whether the same policy, with no retraining, would keep walking on a real vessel rather than a lab platform.
A dry run on a moving floor
Before the ship, we wanted a controlled version of the same problem. On March 24 we put Digit on Georgia Tech’s CAREN system, a six-degree-of-freedom motion platform with a treadmill deck, and ran the L2T policy while the platform pitched and rolled underneath it. The robot hangs from an overhead safety tether that only takes load if it falls.
The platform let us dial in disturbances we could not control on the ship, and it told us the policy would at least tolerate a tilting floor before we committed to the trip.
Setup
We brought the same Digit and the same rolling safety gantry we use on campus. The gantry follows the robot and is there to catch a fall, not to hold the robot up. The policy itself was unchanged from our other hardware experiments: no retraining, no ship-specific tuning.
What happened
The policy walked. We started with passes up and down the bay while the Stiletto was tied up at the pier, then kept running once she got underway. Around noon a crew member held up a phone GPS at the start of a run so the video would record the ship’s speed: 11 mph (about 10 knots) for the first underway runs and 25 mph (about 22 knots) for the last one. The clips below are the full recordings, not highlights. Digit kept walking through turns, stops, and restarts on the moving steel deck. The gait looked the same as it does on campus concrete, which is the point of the privileged-learning setup: the student never sees the terrain directly, so it has no reason to behave differently as long as its proprioceptive history stays in distribution.
At the pier
Underway
Takeaways
Two things stood out. First, zero-shot transfer held up in an environment we had never modeled, including a deck moving at 20-plus knots, which is encouraging for the sample-efficiency argument behind L2T: the policy did not need extra data to handle a new floor. Second, the logistics of a field test dominate the day. Moving a humanoid, a gantry, and a control station onto a ship and running safely in a crowded bay takes far longer than the walking itself, and that is worth planning for before the next trip.
Thanks to the Navy crew who hosted us and to everyone who helped carry gear on and off the ship.