Chuguang Leap raises a seed of 'several hundred million yuan' with Xiaomi and Baidu Ventures

Shanghai Chuguang Leap Intelligent Technology has closed a seed round worth "several hundred million yuan", the company says, with Chentao Capital, Xiaomi strategic investment, Baidu Ventures and Xinhe Capital among the backers. None of the investors is described as lead, and the exact figure was not disclosed. A business-registry change on Oct 12 added Xiaomi's Hanxing Venture Investment and a Baidu-linked fund, Sanya Baichuan Zhixin, as shareholders and raised registered capital from 500,000 yuan to 600,000 yuan.

The seed prices a method and a team, and the benchmark result that gave the company its first headline did not last. Its Simate-beta model, a VLA system, took first place on the RoboDojo leaderboard on Sept 23 with an average score of 33.95 and a 27.96% success rate, ahead of GPT-6-Astra. In the Oct 9 snapshot of the same board, Robotera's VPP2, a world-action model, led at 39.26 and 32.26%, and PhysicalRSI (36.27 and 31.38%) and Awomo-0.5 (35.34 and 29.64%) were ahead of Simate-beta as well, leaving it at best fourth. Its first-place run had ended by Oct 9 at the latest.

Recursive self-improvement is the stated bet

What the company pitches is "Physical RSI", short for recursive self-improvement of physical intelligence. Its AutoResearch system runs a loop that proposes hypotheses, plans experiments and then executes, validates and analyses them automatically, with human researchers setting direction and constraints, the company says. Researchers from MIT, Tsinghua, Peking University and HKUST have joined internal testing of the platform, according to the company.

Driving's end-to-end stack, moved to arms

Founder Zhang Ying was a core technical lead of Horizon Robotics' HSD driver-assistance system, where he led three generations of end-to-end driving systems: HD-map-based, map-free and vision-only one-stage end-to-end. He left Horizon in 2026. His stated thesis is that the end-to-end "perception–reasoning–execution" method proven in autonomous driving generalizes to manipulation in unstructured environments, which puts the young company in the embodied-AI field. The team includes Zhan Fangneng, an HKUST assistant professor who runs the World Mind Lab, and Jimazeyu, a UC San Diego graduate and student of Xiaolong Wang who was a founding member of ARI.

The opposing read is that a benchmark rank is a weak proxy at this stage, and that a seed cheque for an ex-Horizon end-to-end lead buys driving-scale engineering rather than a durable RoboDojo position.

Several numbers already sit ahead of Simate-beta's 33.95: VPP2's 39.26, PhysicalRSI's 36.27 and Awomo-0.5's 35.34 in the Oct 9 snapshot, plus the 44.48 that Einsia AI's Oct 13 EmbodiedEvo paper self-reports on RoboDojo-Sim as surpassing all officially verified results. The next RoboDojo snapshot is the first checkpoint at which Chuguang Leap can show its own research loop moving that number rather than a single hand-built model.