Big thing
Tesla's Robotaxis swarm the Austin F1 Grand Prix
Tesla stretched its Austin geofence a few days earlier to take in Circuit of the Americas, and on race day more than 50 Cybercabs were seen dropping off fans (@SawyerMerritt). Spotters logged a record 1,250 Cybercab sightings in Austin on qualifying day, 1,159 of them driverless (@RobotaxiRadar). After the race, a Cybercab from the track to downtown cost $76.07 with a wait of about 40 minutes, against $126.98 and 12 minutes for Uber Comfort (@RobotaxiRadar).
Celebration, led by Tesla itself: "Swarm" (@robotaxi, 117K) and "Robotaxis as far as the eye can see! 🤩" from its AI head @aelluswamy. Fans noted that no other robotaxi reached the track: "Tesla is the only AV operator that can take riders to and from the F1 race in Austin, Texas this weekend" (@OverlyTrev). @DriveTeslaca: "This must have been a great stress test for @robotaxi, and it seems like it passed with flying colours."
Key research
MEMI: robots that work out what to remember nearly solve RoboMME
Yuejiang Liu and colleagues, with Yilun Du and Chelsea Finn, have models reverse-engineer what a robot should remember instead of being told. On RoboMME's 16 memory-dependent tasks, their Astra-Plan setup goes from 92.4% to 99.4% and the end-to-end Astra-Act from 88.3% to 97.0%, and even Claude Haiku climbs from 73.9% to 90.8%, which they call human level (@liu_yuejiang). Code, prompts and logs are out. It matters because the gain held across all four reasoning models they tried, so better memory, not a bigger model, closed the gap.
FOMI: a scissor-shaped tool that lets you feel what the gripper touches while you record demos
A team from HKUST, HKU and Robbyant kept UMI's handheld, robot-free data collection, but made each handle, gear and link one rigid piece, so contact forces travel back to the hand with no sensors or motors. Three first-time operators collected 17.8 successful demos per 10 minutes against 13.8 with UMI, and π0.5 fine-tuned on about an hour of FOMI data per task averaged 78.4% on four contact-rich tasks, like unzipping pouches and washing dishes (@YinghaoXu1). It matters because the operator feels each contact and fixes it on the spot, so demos get better without recording any force.