Robots that plan live, measured in the open

• Any room: the model never sees the room's position frame, only the gantry's height and each line's direction and length from the 4 calibrated mount points. Trained on a new random room every trajectory. A playroom-only version had binned 5/36 in the bedroom.
• Ties the physics simulator: 20 mm and 0.8–0.9° after 2 s of open-loop prediction in unseen rooms, as MuJoCo in the same room (which knows the room but not the latency, the miscalibration or the payload).
• Against the teleop route (naavox, as published): 5,984 teleop episodes (35 h) and 56 models from ACT (52 M params) to pi0.5 (4.1 B) trained on 4090s, H100s and H200s; on the same laptop one decision costs 87 ms (ACT) to 3.8 s (pi0.5, est.), ours 35 ms on the CPU.
• Real rooms from their recordings: anchors, named places and the floor view imported from naavox/playroom-sep29-2 and bedroom-sep6. 3D rooms: the motors sit in the corners, so a room builds the robot; furniture is only in the mission's costs.

• The giver opens only when the taker's servos say it holds: jaws stalled on the pen, load felt, for 0.4 s. A release is an event on the other hand's sensors, never a clock.
• The taker aims where the giver's body says the pen is: the giver's joints × where the pen sits in its jaws. The one camera sees the pen as a line and corrects that belief (direction and depth).
• The pinch feels the pen's thickness: pinched on its curve instead of across its 11 mm diameter, the jaws close further. "Too thin" means re-grasp lower before it slips.
• Each wrist camera faces away from the other hand: it sticks out 7–10 cm along the pen, and the grip spots are 58 mm apart.
• The reverse hand-over is the same logic with the roles swapped.

Compute question: the planner was born on a Jetson-equipped drone that saw no images. How much does vision add? A survey mission switches the camera on only while hovering: take off, hover at A and look for the red mailbox, fly blind to a stand-off point, hover and look again, land.

• Events, never clocks: seen → approach → close → lift; "closed on nothing", "jaw blocked", "slipped" are read from the servo's angle and current.
• Sense of grip: a gentle hold, slip felt by the gripper servo, squeeze harder only when needed. Same loop grasps a screwdriver.
• The table measured by touch: the arm probes the desk; the model read it 14.5 mm too low.
• Talk to it: a sentence → the arm's own skills, in a simulated pen / highlighter / cup world.

The glass is found by shape (depth → voxel clusters → radius profile matched to a catalogue), never from ground truth. Each fix is logged with its measured effect, including the rejected ones.

On the real arm SmolVLA aligned on the pen but often slipped or landed offset, and nothing in it retries. That is why armpilot exists: explicit closed-loop seeking with costs you can read and edit. The twin lives on as armpilot's world.

• T0 walk / stand: 2.3 cm drift standing 10 s, walks 0.44 m/s, no falls (66 min on an L4, ~$0.90).
• T1 hold a staff: stands, walks and turns with it, but squeezes: 0.9–1.3 kW of heat from the wrists. Nothing in the reward taught it to relax.
• T3 carry a real box (from PRISM's motion clips): picks up, carries, sets down; box 2.0 cm from its reference. Again 1.1 kW at the wrists.
• T5 one generalist for box, ball, barrel, bin: within ~1 cm of the specialists, and carries the barrel its specialist dropped.
• T6 our video → motion: falls at 10%. The lifted reference slides its feet 53 cm/s and puts the box inside the body 84% of frames: plausible, not physical.
• T7 student (no reference clip): falls at ~25%, at the pick. It cannot tell when to crouch.
• Doors: motion capture → door rebuilt from the hand's arc → G1 IK with the scene: 16 of 20 door variants accepted.
• Maps: a real-apartment scan (ReplicaCAD) in MuJoCo; the walking policy goes 8 m around furniture in 18 s; "pick the box on A, drop it on B" planned into skills.

Field → tractor and 3 t trailer → 1 m² reception pad and auger → membrane press → 4 receiving tanks + 24 filtered tanks, with gendered hose couplings and pumps. Cameras include an arm's-eye view of the vine row and a humanoid's eye height in the working aisle: the environment future policies have to work in.

Return to base (~0.15 m final distance on the nominal domain), terrain gates in maps the drone cannot see, ball drops into canyons, craters and between pillars, drop-and-return missions, and a real-time MuJoCo version. The MPPI, task interface and "spaghetti" plan drawing here are what armpilot, barpilot and Hedwig were ported from.