Gaode upgrades ABot with five embodied AI models and claims SOTA on 17 benchmarks

新智元 · wechat · 2026-07-22

The embodied AI race shifts from single skills to full systems

This long-form piece argues that the embodied AI race is moving from “can the robot move?” to “can it reliably do work in the real world?”. It uses Tesla Optimus and Figure as examples of a broader shift toward Physical AI, where robots are expected to perceive, reason, remember, and act in open-ended environments.

The article’s main case study is Gaode’s ABot stack, which it says was upgraded with five new models: N1, M0.5, ER, AgentOS, and C0. The stack is presented as a three-layer loop:

Embodied brain: ER + AgentOS handle perception, reasoning, task planning, and memory.
Motion dual-core: N1 handles navigation; M0.5 handles manipulation.
Training/simulation: ABot-World and ABot-Earth provide simulation and data feedback.

It claims the system reached SOTA on 17 benchmarks and reports strong results for:

N1: open-world navigation with a reported 92.9% success rate on city-scale long-horizon tasks.
M0.5: decoupled mobile manipulation, with a reported 20.4% gain over prior SOTA on RoboCasa-365.
ER / AgentOS: unified reasoning, long-term memory, and plug-in skills for heterogeneous robots.
C0: end-to-end motion validation on Gaode’s quadruped robot “Tutu”.

The broader argument is that embodied AI’s next phase is not isolated component gains, but a closed-loop system where simulation, physical interaction, and memory reinforce each other.

Original post →

Gaode upgrades ABot with five embodied AI models and claims SOTA on 17 benchmarks

More from Companies & People