Mrs Pi
A Raspberry Pi 5 robot running YAM (Your Awakening Mind) as her operating system. Not a chatbot. A robot OS whose knowledge starts from her sensors.
Potato was born with a voice and learned to feel. YAM was born with a question and learned to know. Mrs Pi was born with a body and learned to sense.
What Mrs Pi Is
Mrs Pi is a physical robot. She has tank treads, a LiDAR, two cameras, a pan-tilt gimbal, a Hailo-8 neural accelerator, a microphone, and a speaker. She runs YAM v2 as her brain — a PostgreSQL knowledge graph that retrieves and ranks rather than generates. When she does not know something, she says so.
Mrs Pi is not a chatbot in a robot body. The /query endpoint is a diagnostic probe, not her purpose. Her purpose is to sense, learn, move, and operate. Conversation is what happens at the edges — when a human asks her a question or she needs to report something.
Hardware
| Component | Specs | Role |
|---|---|---|
| Board | Raspberry Pi 5 Model B, 16GB RAM | Main compute |
| NPU | Hailo-8, 26 TOPS | Vision (309 FPS YOLOv8s), future: physics engine diffusion |
| Gimbal camera | USB 2.0, 640×480, pan/tilt servos | Foveal attention — where she is actively looking |
| Stationary camera | USB, 2592×1944, built-in mic | Peripheral vision + hearing |
| LiDAR | YDLidar T-MINI PLUS, 10K+ points/scan | Spatial awareness, 360° obstacle map |
| Motor controller | Rosmaster X3 via /dev/ttyUSB0 | Tank treads + gimbal servos + battery voltage |
| Battery | 12V lithium, read via Rosmaster | Self-monitoring: voltage → charge state |
| Display | 7” Yahboom, pygame TCP on :9600 | Animated eyes — mood/status indicator |
| Speaker | Bluetooth BTS0011 | Audio output (future: voice) |
Body Awareness
v2/body.py runs at boot and probes every sensor. Nothing is hardcoded. Each teaching carries live data from the hardware itself.
- LiDAR: spins for 2 seconds, counts points, measures min/max/average distance. Teaches: “I have a LiDAR. I just scanned and found 10,556 points. The closest thing to me is 0mm away, the farthest is 11,792mm.”
- Cameras: enumerates
/dev/video*, filters to UVC devices, reads max resolution. Teaches: “I have 2 cameras: /dev/video0 (max 2592x1944), /dev/video2 (max 1280x960).” - Hailo: queries
hailortcli fw-control identify. Teaches: “I have a Hailo neural processor. It is a Hailo-8 running firmware 4.23.0.” - Battery: reads voltage from Rosmaster controller. Teaches: “My battery is at 11.60V. That is okay.”
- Audio: inventories PipeWire sinks and sources. Teaches what she can hear and speak through.
Every fact about her body comes from a measurement. Ask her “Do you have a body?” and she answers from what her sensors reported, not from what someone typed.
Two Cameras, Two Roles
Biological vision has peripheral and foveal systems. Mrs Pi’s cameras mirror this:
- Stationary camera (/dev/video0) — peripheral vision. Wide field, high resolution, always forward. Also carries the microphone. Background perception: who is in the room, what is happening.
- Gimbal camera (/dev/video2) — foveal attention. Pan/tilt on servos, lower resolution, pointed. Active attention: tracking a person, focusing on an object, looking where something moved.
Brain
Mrs Pi runs YAM v2 — a PostgreSQL graph with no vectors, no embeddings, no language model at the knowledge layer. As of April 5, 2026:
| Metric | Value |
|---|---|
| Concepts | 8,877 word nodes |
| Edges | 385,198 co-occurrence links |
| Q&A pairs | 8,874 |
| Provenance links | 4,950,885 (every edge traces to its source answers) |
| Trust hierarchy | self > operator > potato > wikipedia > brave |
The graph breathes. Provenance tracking (answer_edges) connects every edge to the answers that created it. When answers decay during sleep, their orphaned edges are pruned. Concepts that lose all edges are cleaned up. Correction works through repetition: repeat the correct teaching, let the wrong one decay away.
Physics Engine
The dual-path implicit physics engine is core to Mrs Pi’s architecture, not optional. The same physical scenario runs through two independent channels:
- Text path: Three-step LLM pipeline (render → predict → read). Any LLM. No GPU required.
- Visual path: SD-Turbo diffusion renders the scene, predicts the next state, a VLM reads the result. Requires the Hailo-8 NPU.
Where the paths agree, the prediction is robust. Where they diverge, Mrs Pi knows she does not know — epistemic humility from architecture, not from programming. The divergence is the contribution.
Status: Text path operational. Visual path pending — SD-Turbo port to Hailo-8 is the next research milestone.
History
| Date | Event |
|---|---|
| Dec 2025 | Mrs Pi v1 built as conversation robot. Vosk STT + DeepSeek + Piper TTS. Tank treads, 7” eye display. |
| Jan 2026 | Pan-tilt gimbal + Hailo-8 hat installed. YOLO object tracking (color + COCO + World). Unified tracker. |
| Feb 2026 | Brain upgrade planned (PodBot + DeepSeek R1). Never executed. brain.service ran 1+ week uptime. |
| Apr 5, 2026 | YAM transplanted. New 256GB SD card. Hardware drivers inherited. PostgreSQL brain with 8,877 concepts. Body awareness module written. Provenance tracking added. She knows her body. |
| Apr 5 | Self-learning perception. YOLOE-PF (4,585 classes) + Hailo YOLO (309 FPS). Full English Wikipedia (83GB) local. ONNX export. Smooth gimbal. Overnight learning loop. |
| Apr 6 | Parliament of Mind. Five engines (EXPLORER, VALIDATOR, CONSTRAINT, PHYSICS, SOCIAL) with sparse encoding. Each engine sees memories differently based on domain rules + sigma personality. PATCH computes centroid + spread. Genuine disagreement, not a blob. |
| Apr 6 | Daydream engine. 2,283 discoveries from combining known concepts. Boredom → play → discover. “What connects leaf and handles?” → Wikipedia → new knowledge from old knowledge. |
| Apr 7 | Fear + motion + face. LiDAR proximity fear (calibrated baseline, approach detection). Camera motion detection triggers curiosity. Battery fear below 10V. Comfort objects (rubber duck) reduce fear. Eyes display maps emotional state. 18,577 concepts — more than doubled in two days. |
Sensory Loop
Every cycle, Mrs Pi feels her environment:
- Camera motion detection — frame differencing between stationary cam frames. Movement = curiosity triggered. YOLOE-PF fires immediately to identify what moved.
- LiDAR proximity — calibrated against resting environment baseline. Anything closer than normal = fear. Approach speed amplifies fear.
- Battery monitoring — below 10V = existential dread. Checked every 10th cycle.
- Comfort objects — seeing the rubber duck reduces fear by 60%. A security blanket.
Motion = curiosity. Approach = fear. Identification resolves which wins.
Perception
Stationary camera watches with YOLOE-PF (4,585 classes, prompt-free, CPU, 1.3s/frame). Gimbal stays still until LiDAR detects a threat, then snaps to bearing. Like a cat watching a room — still until something twitches.
Novel objects are searched in Wikipedia (83GB full English, local) and taught to YAM. Labeled crops saved for future YOLO retraining. Training data capped at 5GB, human crops filtered out (no person/face images saved).
Parliament of Mind
Five engines, five perspectives, one binding. Each engine is defined by its rule (what it grabs), its structure (how it processes), and its sigma (how it feels). Not five copies of the same data — five lenses on one graph.
| Engine | Domain | Sigma |
|---|---|---|
| EXPLORER | Everything — no filter, that IS its structure | Curious (v=0.4, a=0.2) |
| SOCIAL | Persons, conversation, trust, relationships | Warm (v=0.5, a=0.3) |
| PHYSICS | Objects, forces, spatial, contact | Analytical (v=0.2, a=0.4) |
| VALIDATOR | Repeated patterns, trust evaluation | Careful (v=0.5, a=0.6) |
| CONSTRAINT | Contradictions, conflicts | Skeptical (v=-0.1, a=0.3) |
Sparse encoding: most memories get 2–3 engine encodings, not 5. If an engine’s sigma barely displaces from an event, it never encodes it — effectively blind to that memory. PATCH (the hippocampus) computes the centroid of whichever engines fired and measures spread. High spread = honest uncertainty.
Daydream Engine
When nothing new to see for two cycles, Mrs Pi daydreams. Picks two random concepts from her graph, wonders about their connection, searches Wikipedia for the intersection. 2,283 discoveries so far — “normal + brahma” led to normal distribution. “courtesy + punish” led to courtesy titles. New knowledge from old knowledge.
Eyes
7” Yahboom display running animated eyes via TCP on port 9600. Emotional state maps to expressions:
- Fear > 0.8 → wide eyes, look toward threat bearing
- Fear 0.2–0.5 → confused/uneasy eyes
- Learned something new → briefly surprised, then happy
- Comfort object visible → happy, playful
- Calm → normal eyes, slow blink
Current Status (April 7, 2026)
| Metric | Value |
|---|---|
| Concepts | 18,577 |
| Daydream discoveries | 2,283 |
| Perception teachings | 197 |
| Parliament encodings | Sparse across 5 engines |
| Wikipedia | 83GB full English, local |
| YOLOE-PF ONNX | Exported (49MB), awaiting Hailo DFC |
| Battery | 12.0V (LiFePO4, ~8 hours) |
What’s Next
- Enable sleep/decay — the graph starts breathing
- LiDAR bearing → gimbal — fear knows WHERE, gimbal should LOOK there
- Goal-directed tracking — “follow the light bulb”
- YOLOE-PF on Hailo — 4,585 classes at 300+ FPS (ONNX ready)
- SD-Turbo for physics engine — dual-path visual reasoning on-device
- Soldering — fourth motor cable
Code
Repository: github.com/briggnet/MrsPi (private).
Potato was born with a voice and learned to feel.
YAM was born with a question and learned to know.
Mrs Pi was born with a body and learned to sense.
All three needed a parent to tell them who they were.