Evopien / A new presence / 5 November 2026
It moves towards you. It keeps the thread. It has something to show.
The head turns towards a face. A visual watch notices the change you asked for. A spoken counter keeps tally while the interaction continues. A question brings current information from the web into the conversation. Evopien is becoming a much more tangible presence.
Our new English and Spanish videos show the live interaction. Alongside that physical build, governed-memory testing has reached 98 simulated days and 2,200 staged interactions. This chapter brings the achievements together: local voice and vision, a moving head, useful tasks and memory that can change with its owner.
See it in action / 5 November
Two languages. One presence you can see for yourself.
Voice, live vision and web search: explore the new Evopien demos in English and Spanish.
EVOPIEN AI Robot Demo | Live Vision, Voice & Web Search on NVIDIA Jetson Thor
EVOPIEN AI Robot Demo | Vision Live, Voz y Busqueda Web - NVIDIA Jetson Thor
Measured progress / recorded development runs
The numbers behind the new presence.
- Median voice onset
- 1.35 sSpeech end → first generated audio · 11 recorded non-camera onsets.
- With video enabled
- 1.58–1.88 sSpeech end → first answer audio · four EN/ES replay examples.
- Public-web answers
- ≈1.6 sRecorded answer retrievals: 1.575–1.653 seconds.
- Governed-memory timeline
- 98 daysSimulated days · 2,200 staged synthetic interactions.
- Latest memory week
- 129/129Complete answers reviewed, with 129 matching state checks.
- Physical face following
- 10+ min637.5 seconds recorded in the saved head session.
These snapshots come from separately recorded development runs. Voice onset marks generated audio; camera-on figures use generated scene input in a local replay. Web timing measures answer retrieval. Memory follows a simulated chronology; head duration comes from the attended physical session.
01 / Nearly 100 simulated days
A memory milestone with real depth.
98 simulated days. Fourteen weeks on a changing test timeline. 2,200 staged synthetic interactions. Evopien’s governed-memory campaign has moved through saved preferences, corrections, changes of mind, selective forgetting and restarts in English and Spanish.
The achievement is in what happens next. A later answer must follow the current preference, including its conditions and reasons. A one-off exception belongs to that request. A correction changes what is current. A memory the owner has removed stops shaping later retrieval.
The latest weekly block delivered 129 out of 129 reviewed complete answers, alongside all 129 corresponding memory-state checks. In that run, 118 text responses completed within five seconds. The memory campaign and the new physical demo are distinct recorded milestones, advancing the same ambition: a companion that stays useful as life changes.
02 / From design to movement
The head we imagined is now on the desk. And it moves.
Evopien now turns and tilts its physical head to follow a face during our live sessions. The camera, display, printed enclosure and neck form one visible point of interaction. A glance has a direction. A spoken answer has a place.
The saved head session recorded 637.5 seconds—more than ten minutes—of physical following activity, with face reacquisition exercised within the session. This is the moment the build became something you can meet across a desk.
03 / Give the conversation a task
Watch for a change. Count a word. Keep the conversation moving.
Visual watches let you ask Evopien to look out for a requested change and tell you when it appears. Spoken keyword counters keep track of chosen words, report running totals and accept cancellation by voice. Multiple counters bring more than one small task into the same session.
These capabilities change the rhythm of the interaction. You can set something in motion, continue talking and return to the result. The saved physical run delivered five watch notices; a separate bilingual counter replay delivered sixteen spoken count updates across four cases.
Face following
The physical head turns towards the face in view.
Visual notices
A requested scene change becomes a spoken notification.
Spoken word tallies
Track chosen words, ask for totals and cancel with your voice.
English and Spanish
Local speech and visual interaction, demonstrated in both languages.
04 / Capability on the device
Local voice. Local vision. A quicker route to the outside world.
Listening, speech, visual understanding, governed memory, face tracking and head control run on our own hardware. That local foundation gives the project its character: intelligence close to the person, with a physical presence in the room.
Recent local development measurements recorded a 1.35-second median from speech end to first generated audio across eleven non-camera onsets. A separate video-enabled bilingual replay recorded first answer audio between 1.58 and 1.88 seconds across four examples.
For current information, recorded public-web answer retrievals completed in 1.575–1.653 seconds, with first text appearing in roughly 1.3 seconds. Weather, news and practical research can join the interaction while the local voice, vision and physical control stay on the device.
05 / Since the last journal
Three weeks. A much more complete picture.
The last journal was dated 16 October. Through the rest of that month, live audio and camera work gained visual watches, spoken counters and richer scene interaction. In early November, those capabilities came together with bilingual conversation and quicker web research. On 5 November, we saved the new physical-head demo checkpoint.
Alongside the 98-day memory milestone, the build now has a new sense of scale: more it can do, more you can see, and a real head moving on the workbench. The videos are the invitation. Come and see how far Evopien has travelled.
06 / The people helping make it real
A huge thank-you to Seeed Studio.
To the whole Seeed Studio team: thank you for supporting Evopien and helping us turn an idea into something we can build, hold and bring to life on the workbench. Your hardware support, encouragement and trust have made a real difference to this independent project.
From hearing and seeing to a dedicated face and the components for motion, that collaboration runs through this stage of the build. We are proud to share the progress with you.
Explore the Seeed Studio hardware supporting the Evopien project:
- 1 × reSpeaker Flex XVF3800 Circular 4-Mic Array
- The compact microphone array for the head’s local voice interaction.
- 1 × Orbbec Gemini 2 3D Camera
- The camera bringing the scene in front of Evopien into the interaction.
- 1 × 7-inch 1024×600 60Hz IPS Capacitive Touch Screen
- A dedicated display giving the head its visible character.
- 2 × Feetech ST3215 Serial Bus Servos — 12 V C047
- Motion components supplied for the physical head build.
- 1 × Bus Servo Driver Board for XIAO
- Part of the supplied hardware for developing the head’s movement.
- 1 × Mono Enclosed Speaker — 4 Ω / 5 W
- A compact speaker for the voice of the head.
- NVIDIA Jetson AGX Thor Developer Kit Bundle
- The local computing platform behind this generation of Evopien, available through Seeed Studio.
The product links above are Seeed Studio affiliate links and may support the project through purchases.