A 7B open model takes first place on RoboLab
Black Forest Labs says its new 7B open-weights robot model beats a 16B rival on RoboLab-120, 42.9% to 36.8%, while running up to 3.95x faster. Plus two humanoids playing soccer with nothing but onboard sensing.
Shown this week
Black Forest Labs releases FLUX 3 Action, a 7B open robot model
Black Forest Labs released FLUX 3 Action, an open weights 7B world action model (predicts future video and robot actions together) it says takes first place on the RoboLab benchmark.
F3A is SOTA without the WAM latency: Beats Cosmos 3 Nano (16B) on RoboLab-120 (42.9% vs. 36.8%) while running at less than half the parameter count (7B) and up to 3.95x faster.
MIT takes humanoid soccer outdoors with no instrumentation
Pulkit Agrawal's group put humanoid soccer on real fields around people, with no instrumentation and no controlled environment. He says the policy was trained in simulation and deployed with no post-training, doing dribbling, passing, turning and scoring. The work is led by Avi, Khai and Nolan Fey with Martin Peticco, John Marangola and Venki.
Trained in simulation. Deployed in reality. No post-training. Just pre-training.
Unitree G1 dribbles through slalom courses on onboard LiDAR alone
Chuye Hong showed a Unitree G1 keeping control of a soccer ball through slalom courses and large direction changes, not just straight-line pushing. Ball perception is fully onboard with LiDAR, no mocap and no external tracking, and Hong says the policy trained in simulation with a human-motion prior and transferred zero-shot to hardware. Chris Paxton called it really impressive.
Ball perception is fully onboard with LiDAR — no mocap, no external tracking.