Building in the open
This is where the proof lives: the efficiency and deployability findings behind the bet that intelligence belongs on-device.
We trained a competitive Vision-Language-Action policy on a single RTX 5090 in a home office — about 7.5 hours, under $10 of electricity, and 81.8% average success on LIBERO. The weights are fully open under Apache-2.0.
Read moreLarge VLA models are powerful, but real robots run under latency, memory, power, and cost constraints. This note explains why useful capability per parameter matters for deployable robotics.
In preparationA deployment-first architecture direction: compact multimodal fusion, robot-state conditioning, efficient action decoding, and hardware-aware inference.
In preparationParameter count alone is not enough. We discuss how to evaluate latency, memory footprint, action success, robustness, fine-tuning cost, and hardware target.
In preparationNotes on building an evaluation pipeline that starts in simulation and moves toward real robot deployment without losing sight of practical constraints.
In preparation