KDDI Research Tops RoboChallenge Real-Robot AI Benchmark With Blended VLA Models
KDDI Research, the research arm of telecoms operator KDDI, says it has taken first place worldwide on Table 30 V2, the newest evaluation track of RoboChallenge, a global benchmark that tests robot AI on physical machines rather than in simulation. The company announced the result on October 5 and noted that the ranking was current as of that date; because many teams test on the track every day, success rates and positions change daily.
RoboChallenge describes itself as the largest platform of its kind, and more than 100,000 real-robot tests have been run on it so far, with entrants that include major Chinese technology companies and leading humanoid developers. Table 30 V2 uses four types of robots and 30 real-world tasks focused on hands-on manipulation, including hard cases such as coordinating two arms and handling soft, deformable objects.
KDDI Research credits two techniques. The first blends knowledge from several vision-language-action (VLA) models, which turn camera input and natural-language instructions into robot motions. Instead of training a model to copy a stronger one exactly, its method follows the stronger model closely only at the moments that decide whether a task succeeds, such as how firmly to grip or when to switch hands, and allows flexibility elsewhere. The second uses a world model, an AI that predicts physical cause and effect, to forecast how each candidate plan would play out, pick the one most likely to succeed and adjust the robot’s speed. Among the models it drew on was DM0.5, an open-source VLA model released by Dexmal.
What it means in practice
Benchmarks on real hardware are harder to game than simulated ones, but first place on a moving leaderboard is a snapshot. The more telling step is where the technology goes next. KDDI Research says it will apply the method to its project under GENIAC, the generative AI programme run by the Ministry of Economy, Trade and Industry and NEDO, which aims to build robot foundation models that can take over frontline work in retail and logistics. Professor Jiang Liu of Waseda University supported preliminary tests on physical robots and model tuning.
The result adds a telecoms group to Japan’s physical AI push, alongside the Hitachi and Fanuc factory partnership, national machine IDs meant to feed physical AI training and Kawasaki’s new Home Leo and welding robots.
What to watch: whether KDDI Research holds its position as other teams submit new runs, whether it publishes technical details, and when GENIAC-backed robot models reach real stores and warehouses, as well as manufacturing and facility maintenance, the other fields the company names.