Overview
During the holiday, on-device embodied intelligence broke through: YidaTech launched an “Edge AI & Embodied Robot” solution that moves AI compute and decision control from the cloud onto the robot body. End-to-end latency is under 10 milliseconds (about 90% lower than cloud schemes), enabling over 8 hours of autonomous operation even in weak or no-network environments.
At WAIC 2026, ModelBest released the MiniCPM-Robot on-device series: the 0.9B RobotTrack is the first locally deployable offline tracking model, running stably on Unitree Go2’s native compute at ~180ms latency; the 1.5B RobotManip supports up to one minute of embodied native memory and can replan tasks after interruption, all via local inference.
Background & Interpretation
1. Technical Background and Evolution
Traditional robots rely on cloud inference, which fails in factories, underground or disaster sites due to latency and instability. Moving compute onboard is the key step from “demo” to “real work”. Small on-device models (<2B params), via distillation and quantization, run tracking, navigation and manipulation loops on limited silicon.
2. Core Drivers and Mechanisms
Drivers: (1) edge chips and NPUs now run multimodal models; (2) hard reliability needs in no/weak-network sites (mines, warehouses, inspection); (3) data privacy and real-time demands favor local inference. Key tech: model lightweighting, low-latency pipelines, local long-horizon memory.
3. Market Structure and Industrial Chain Effects
The solution is already used in casting defect inspection, green-mining autonomous vehicles and weak-network warehouse sorting, and was selected for a CIFTIS case. A new collaboration forms among edge AI chips, robot OEMs and system integrators. Versus cloud-dependent designs, on-device cuts bandwidth/cloud cost and boosts availability.
Implications & Outlook
Near term, industrial inspection, guiding and warehousing will scale first; medium term, a hybrid of on-device autonomy plus cloud coordination may dominate, upgrading robots from “controlled devices” to “autonomous colleagues”.
Takeaways for Industry Participants
- Robot makers: make on-device inference a core differentiator.
- Chip/algorithm firms: deepen <2B multimodal models and low-latency deployment.
- Industry users: prioritize on-device reliability in critical weak-network scenarios.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.