Embodied AI Model Tops Benchmark as Industry Faces Reality Check

RoboDojo was created by the University of Hong Kong, UC Berkeley, Tsinghua University and nearly 20 institutions, serving as a unified test for embodied models. It covers 42 simulation tasks and 18 real-robot tasks across five dimensions: generalization, memory, fine manipulation, long-horizon execution and open semantic understanding. As of July, head models scored only 8.80% on simulation and 12.8% on real robots, compared with 76.03% and 100% for human experts. The worst category, open semantic tasks, had a 1.67% success rate. DM0.5's Memory score was 47.74, more than triple the second-place 13.37, and it passed the Cover Blocks memory test with 100% across three random seeds.

The model uses a 4-billion-parameter VLM backbone plus a 680-million Action Expert, with a context abstraction layer that natively supports up to 60 seconds of memory. Unlike most VLA (vision-language-action) models that react frame-by-frame, DM0.5 compresses history during pretraining, and its "embodied chain-of-thought" forces reasoning before action. Engineering optimizations cut core inference latency from 534 ms to 57.49 ms, a 9.29x speedup, while maintaining accuracy. Official demos show it assembling micro building blocks, peeling cucumbers with force control, and sorting parcels at 3-second intervals with 99% accuracy.

The result highlights a wider industry gap. At the 2026 World Robot Conference, a forum on "The Distance from World Models to Serving Humans" brought together eight entrepreneurs in model, data, hardware and simulation. Chaired by Wang Jian, director of Zhejiang Lab and founder of Alibaba Cloud, the panel discussed the limitations of current VLA models, which one participant called "imitation learning" that cannot truly understand physical rules. Another described world models as "the pearl in the crown of generative AI." They agreed that data scarcity, edge computing limits and safety mechanisms remain hurdles, and that deployment should start in industry, then expand to commercial scenes before entering homes.

The investment ecosystem is also responding to AI entrepreneurship. In Shenzhen, Delin Capital launched the Delin Cup 2026 AI Startup Competition, offering 2 million yuan in unconditional cash to the champion and a 40-million-yuan, 4,000-square-meter incubation space called Delin Residency, as reported by QbitAI. Up to 15 teams will receive free office, housing, food, computing and hardware lab access, plus potential equity investment of up to 5 million yuan and a demo day with more than 30 top VCs. Applications close Sept. 30, with the finals scheduled for Oct. 15-16.

On the open-source front, at an APEC forum on SME digital transformation, Yang Qian, co-founder and COO of Zibian Robot, said open-source embodied AI models can cut development time dramatically. She cited a team of young developers who built a platform in three days using open-source foundation models, a task that used to take at least six months. The company has open-sourced its end-to-end model WALL-OSS, including weights and training code. SMEs account for about 97% of APEC member economies' businesses, she noted.