Open-source robot brain DM0.5 takes the crown on RoboDojo
The hardest leaderboard in robotics has a new number one — and this time, anyone can download it.
Dexmal's DM0.5, an open-weight vision-language-action model, has taken first place on RoboDojo, according to Dexmal and QbitAI's report on the leaderboard. Its composite score of 24.90 and 19.34 percent average success rate sound modest until you see the field: as of July 2026, models at the top of RoboDojo's simulation board averaged just 8.80 percent success, and the real-robot board 12.8 percent. The benchmark — built by nearly twenty institutions including the University of Hong Kong's multimedia lab, UC Berkeley and Tsinghua — is engineered to be brutal, with long-horizon tasks, deformable objects and active interference, precisely so polished lab demos can't fake general competence.
The decisive edge came from memory. DM0.5 scored 47.74 points on RoboDojo's Memory dimension, far clear of the pack, and aced the Cover Blocks task — where a robot must memorize the hidden positions of colored blocks and later uncover them in order — with a perfect 100 percent across three randomized trials. The model natively holds up to sixty seconds of history inside its 4-billion-parameter vision-language backbone, and that continuity is what lets it learn a new task from a single human demonstration video instead of thousands of scripted repetitions. Most VLA models treat every moment as if it were the first; Dexmal bet that persistent memory is the unlock, and the scoreboard agrees. It's a notable datapoint in the robot-brain race — a different lane from Moqi's agentic-native robot brain running a 15-minute chore loop, which leaned on agentic planning rather than raw recall.
This isn't a one-board wonder, either. Dexmal's published results show DM0.5 at 99.0 percent on LIBERO, 93.6 percent on RoboTwin 2.0's clean setting (versus 82.7 for Pi0.5), and 43 percent success on RoboChallenge Table30 v2 — a physically grounded real-robot suite where Pi0.5 manages 14.3 percent. The team also cut core inference latency from 534 milliseconds to about 57 milliseconds, a 9.3× speedup with essentially no accuracy loss, which is the difference between a demo reel and something that runs on actual hardware.
What makes the result travel is that Dexmal shipped the whole recipe. The weights are on Hugging Face and ModelScope under a Gemma license, the training framework and evaluation workflows are Apache-2.0 on GitHub, and even the data mix is disclosed: fifty thousand hours of teleoperated robot data, a hundred thousand hours of egocentric video, and simulations covering a million square meters of indoor scenes. Open weights are routine in language models; in embodied AI, where physical evaluation is expensive and reproducibility is scarce, publishing the full methodology behind a number-one result is still rare enough to matter. As we noted with NVIDIA's open-source Cosmos 3 world model, the interesting competition in physical AI is quietly shifting toward whoever opens up fastest.
What to watch: whether DM0.5's lead holds on the real-robot board — embodied-AI champions have swapped places within weeks — and whether rival labs answer with full open recipes of their own.
If you had to bet on one capability to decide the robot-brain race — raw dexterity or long-horizon memory — which would you pick? Tell us in the comments.
Sources: Dexmal OpenDM (GitHub)
