Perception, planning and real-time control with safety independent of software.
A manipulation platform where perception and planning run on an application processor while motion control stays deterministic on a real-time core. An independent safety layer can override everything above it.
Perception needs throughput and motion control needs determinism, and they fight for the same machine. Safety behaviour could not depend on a general-purpose scheduler, and mechanical iteration was too slow to absorb late architectural change.
We separated the two concerns behind an explicit contract: the planner issues setpoints on a bounded cycle, the real-time core executes them, and a hardware-enforced safety layer sits outside both. That let perception evolve without ever destabilising motion.
Harmonic-drive joints with encoder feedback and per-axis current sensing, designed against defined load cases with hardware emergency-stop paths.
A perception stack for detection and pose estimation, SLAM for localisation, and a planner running on a bounded cycle.
Real-time motion control on a dedicated MCU, ROS 2 on the application processor, and an operator console with full-fidelity run logging.
Mechanical
AI
Embedded
Apps