Big thing
Tesla confirms 300+ Cybercabs in Austin, 8x in 30 days, and names Dallas next
Yesterday's count came from Texas DMV records (319). Now Tesla says it: "One month after launch, we've scaled to 300+ Cybercabs serving Austin", the first active-fleet number it has confirmed (@SawyerMerritt). On the second anniversary of the We, Robot unveil, Tesla VP of AI Ashok Elluswamy added: "Dallas soon."
Cheered by Tesla fans. @aelluswamy: "We are learning to be on time 😂". @SawyerMerritt: with 2,000+ fans in town for X Takeover, "Cybercab wait times in Austin are still very short." Per @TeslaHype, Musk said there that one injury could get regulators to shut the program down, so safety comes first.
Waymo added about as many cars in Texas in September alone (from about 700 to 1,102, mostly its new Ojai minivans), it just doesn't have a Sawyer Merritt. So Texas already has more than 1,800 robotaxis, close to California (about 2,100), and I bet it passes California before Christmas (about 300 more Cybercabs are already parked in Dallas).
Labs news & launches
1. Ollie: a robot cat built inside its maker's late cat, with Grok as the brain
A custom robotic frame (actuators, cameras, microphones, touch sensing, edge compute) inside the cat's preserved exterior, with motion and routines shaped from years of old iPhone footage. "The hardest part wasn't making him intelligent. It was making him stop moving like a robot."
Stunned questions, then a memecoin. @AmirHadiEng: "did he just use his dead cat to create an AI cat?" Within hours, traders were pushing a $OLLIE token (@tipoisol).
Heat: medium · 671K views, 19 posts
2. A leg-massage robot that is "EXTREMELY aggressive"
Mrwhosetheboss tried a robot that wraps around your leg and moves on its own; a phone app picks the spot (lower or mid thigh) and the pressure (@aiceodev). The maker isn't named.
Jokes. "Robot massage is just a gym membership that finally hits back" (@ItsGoharr). @FrankisTrying: "a massage robot that gets described as extremely aggressive is just a fight you booked an appointment for".
Heat: high · 1.7M views, 8 posts (1.6M on the video)
Key research
WUJI teaches its robot hand to spin a pen in simulation alone
The pen-spinning policy for the 20-joint WUJI Hand 2 was trained with PPO on 4,096 simulated hands in mjlab (built on MuJoCo), following 40 recorded pen motions, then run unchanged on the real hand. WUJI released the code, motion clips, trained models and printable pen files under Apache 2.0. It matters because in-hand manipulation is the hardest dexterity test, and any lab with the hand can now repeat the whole recipe.
SenseNova-RoboRSI nearly doubles a robot benchmark without touching the model
SenseTime's agent harness improves itself from physical task feedback (execute, diagnose, improve, validate, inherit) while GPT-6 Astra's weights stay fixed, lifting its RoboDojo score from 28.97 to 56.83 (50.83% success), with 94.50 on LIBERO-PRO and 64.60 on RoboCasa. Code and a technical report are coming. It matters because the gains come from the harness alone, which is far cheaper to iterate than a robot model.