Test sessions

Test sessions on the robot

Details and numbers are in docs/HANDOFF.md and the results sections of docs/ROADMAP.md.

Date	Where	What	Findings
2026-09-11	Virtual Pepper (NAOqi desktop build)	Every bridge endpoint against real NAOqi	Three bridge bugs fixed (sonar 0.0, missing awareness stimulus, awareness vs head moves)
2026-09-22	Office	First contact: all endpoints, events, camera, microphone, emergency stop, touch reactions, first voice	Five hardware-only bugs fixed; voice 7-12 s to first word, cut to 2.6-3.9 s
2026-09-22	Open public room	Base moves with odometry; model comparison (typed and spoken); speech-to-text recordings	M0 complete; Sonnet 5 medium chosen; recogniser the weak link in noise
2026-09-22	Offline	Speech-to-text comparison on the recordings	Nemotron streaming chosen (6 % vs 37 % word errors)
2026-09-23	Office	Nemotron live; people smoothing and "Around you" line	5/5 spoken prompts heard exactly; people events stable
2026-09-29	Office, then open room	Sonnet 5.5 vs Sonnet 5 (typed and blind spoken); local Qwen3-VL on Alien3; greeting newcomers	Sonnet 5.5 chosen (about equal spoken, faster first words typed); local model 2x faster but describes without looking; greeting passes: picked up at 2.9 m, head turns at once, speech ~2 s later; head reflexes and face tracking only while someone is in view
2026-10-01	Office	0.4.0 checklist: autostart, sharp photos, neutral pose, look-back	Autostart verified after a reboot; photos sharp (400-1411 vs 76-200); two bugs fixed on the robot: an old photo described as the current view, and a paused face tracker pulling the head back before the shot (tracking now held off)
2026-10-08	Office, then open space (solo)	System one in the loop (0.5.0): command router, addressee gate, camera stream, session records, fail-open	Router started 32 actions about 0.25 s after the transcript, one wrong ("turn back" → 180°); handshake rebuilt and detected from the shoulder current; a wave starts a reply about 3.6 s later; camera stream costs the robot 5-22 % of a core; a minute without the local models: everything answered. Fixed on the day (0.5.1): the gate dropped things said to Pepper when alone, turning to face someone, people reports, outage logging. Open: side talk with two people, sentences cut in two (#26), remembering where people were (#27), "come to me" (#24). Details: docs/HANDOFF.md, #22
2026-10-20	Office (solo)	0.6.0 and 0.7.0 on the robot: watchdog, pose, sound direction, pointing, unfinished sentences, reply length, memory by voice, scene notes, initiative	Watchdog, pose and push detection, pointing and look_at, memory with consent and "forget me", shorter replies (16 % of replies four sentences or more) and initiative all worked. Fixed on the day: tablet subtitles never showed, sound direction lost after a bridge restart, two answers ignored by the gate, a held fragment starting a turn, end-of-turn pause 1.0 → 1.4 s. Open: voice direction behind Pepper (#31), a monitor seen as a person (#32), face identity across encounters (#14), come to me in the open space, side talk with two people. Details: docs/HANDOFF.md
2026-10-22	Office, then the open space (with a colleague for a few minutes)	Cold boot with the watchdog, the 1.4 s end-of-turn pause, side talk with two people, voice direction, come to me	The watchdog came up cold (#15 closed). The pause felt a little long, so it is now 1.2 s (#26 closed). Side talk failed: Pepper joined in on all 12 utterances, because both people stood out of view and the "alone" rule answered them (#25). Voice direction was mostly right, but Pepper preferred an old sighting; the voice now wins (#31). Drives worked once the way was clear.
2026-11-05	Office, then the open space (with a colleague for a few minutes)	Bridge 0.8.0, live beamforming, the Cohere second pass, continuous camera recording, a clip-on microphone on Pepper's chest	With Cohere every microphone got nearly every word in the office. In the open space the beamformed head beat the clip-on (13.7 dB against 10.4 dB; about 12 against 8 at 4 m), so the clip-on is not used. The second pass took about 0.08 s from Alien3. Pepper turned its body unasked and the wrong way (fixed offline, not yet rechecked on the robot), and it still joined in on side talk (#25)

Session records

Every robot session is recorded in full (host log, utterance audio, photos, bridge log) and reviewed afterwards; findings become issues. The records hold real people's voices and images, so they stay on the host machine and never go into git or issues. Where each session's records are: docs/HANDOFF.md, "Session records".