Say “come here” and a physical GO2 turns toward your voice, finds you, walks over and sits.

An assistive robot is only useful if it can come to the person who needs it, on a spoken word, without a phone, a map or an internet connection. Come Here is that behavior end to end on real hardware: from a spoken command to a robot sitting beside the caller, with the whole loop running on the robot itself.

It is the behavior the reliability work on this site (HELIX, Phoenix, BlackBoxRS) is meant to make trustworthy.

Hear

A far-field front end on the ReSpeaker array, an energy gate with pre-roll, and faster-whisper base.en on the Jetson CPU, matched against “come here”. Offline, on 57 “come here” utterances recorded on 2026-09-14, the matcher detects 49 with no false wakes on 20 negative utterances; it was tuned on the same recordings.

Localize

One bearing per matched utterance from the array's built-in direction of arrival. Without a confident bearing the robot does not walk.

Turn

Turns in place until odometry yaw has moved by that bearing: a closed-loop turn, not a timed one.

Acquire

YOLO11n person detection, only inside a ±35° gate around the voice bearing; two consecutive fresh detections before any motion. If nobody is in view it asks “Where are you?” once and turns toward the next call.

Approach

Align in place (yaw only), then walk straight (no yaw), with hysteresis and minimum phase times.

Stop

When the person's bounding box fills 82% of the frame height, with backstops for a lost caller, a dead camera, a 20 s limit and a distance budget.

Sit

A final align, one second standing still, sit, and a spoken “Made it.”

The GO2's stock gait cannot combine forward motion and yaw cleanly, so the controller alternates yaw-only align and straight walk phases with hysteresis instead of steering while walking. It stops on the person's bounding-box height, because LiDAR and pinhole distance read long at close range.

Safety is a separate, testable layer. A pure motion gate rejects malformed, non-finite or absurd commands and combined forward-plus-yaw commands, clamps the rest, and stops the robot half a second after commands stop arriving. The e-stop latches in both the bridge and the state machine; releasing it does not resume motion. The bridge refuses to move unless the robot reports the expected motion mode (mcf), and on shutdown it publishes a stop before the process exits. Unit and process tests cover each of these, and the e-stop latch has also fired on the robot during the live sessions.

Seven documented live attempts are listed below, including failures.

2026-09-14
◆Success

Caller at the robot's left, about 90°. Turn, one search turn, align, walk, stop, sit.

2026-09-15
○Failed

Caller at the right. The bearing read +141° and nobody was acquired.

2026-09-15
◆Success

Caller at the right. Turn, walk, stop, sit.

2026-09-15
◆Success

Blind trial, caller position undisclosed. Turn, align, walk, stop, sit, seated about 12.8 s after wake-phrase detection.

2026-09-18
○Failed

Caller at the left. The bearing read +2° (straight ahead), so no turn and nobody in view; the robot did not move.

2026-09-18
◆Success

Caller at the right. Turn, one align turn, walk, stop at 0.83 of frame height, sit, about 12.5 s after wake detection.

2026-09-23
◇Approach completed; sit-complete log not captured

First call read straight ahead and nobody was in view. It asked “Where are you?”, heard a second call, acquired the caller and stopped on the bounding box.

Boot service. On 2026-09-18 a systemd service started the live stack by itself after a cold boot of the Jetson, after one automatic retry. Both 2026-09-18 trials and the 2026-09-23 trial ran on a stack started that way.