Out with Vita
Quick IRL check: pizzeria crowd and the “girl at the car” moment stood out and confirmed people want to engage with robots. It was fun to see Vita in a new environment and we learned a lot. We didn’t capture much, just some photos, but it was enough to validate the vibe. IRL sessions aren’t weekly yet; we’re working toward that as we smooth a few things
On the bench
- Arms and claws are mounted (passive). They look great on camera and in person, and help us validate proportions, mounts, and clearances
- Torso iteration landed with practical tweaks (more mic openings, enlarged camera hole, corrected charging port, stronger head mount and shoulder bases, adjusted current-sensor mount). The silhouette feels closer to “production”
- Print reliability got a bump: quick dry-cabinet for filament, standardizing matte PLA for torso to reduce striations, and general assembly cleanup
Under the hood
We pushed on speed, then chose quality.
- Quantization trial: Merged our LoRA into the 8B model, quantized, and brought it up on TensorRT with NVFP4. We briefly hit about ~120 tokens/s, which was great for raw speed, but the quality regressions were too noticeable for conversation
- Current serving: Rolled back to the standard PyTorch/Transformers stack on our GPU. We’re steady at ~60–70 tokens/s with sub-1s model responses and roughly ~1.5s end-to-end including TTS. This felt best in the real world this week
- Memory & skills: Continued work on long-term memory with local Postgres + RAG, seeding it with past conversations to jump-start context. Also started function-calling tests so Vita can trigger all sorts of actions autonomously
The takeaway: speed is nice, but interaction quality wins. We’ll revisit quantization once we can keep the tone where we want them
What we’re building toward
- V2 thinking: A torso screen for QR codes and playful accessories, exploring Dynamixel-style servos for cleaner wiring and richer control, and making the arms/hands functional. Plan simple grip prototypes and control loops, map power and cable routing, and evaluate servo options and mounting points
- Production mindset: This week wasn’t about piling on features. We focused on simplifications and robustness so we can ship more often: fewer moving parts, faster setup, more consistent behavior. Lots of tiny fixes added up
- More outings: Expect frequent short sessions in varied environments. We’ll start posting clips of Vita interacting unscripted with real people on Youtube, TikTok, Instagram and Base.app in the near future
Next up
- More IRL time
- Keep the current serving path while we improve memory, visual capabilities, autonomous function calls, and stability
- Parallel V2 prototyping on arms, hands, and other ideas we have