Big thing
Toyota will build Waymo robotaxis on its own factory line, starting with Japan
In the next phase of their partnership, Toyota will assemble Waymo-equipped bZ4X electric SUVs directly on its line in Japan (@verge). Testing and validation of the fleet there starts in 2027, with the UK and other Waymo markets to follow (@SawyerMerritt).
Google's @sundarpichai: "こんにちは、日本！Hello Japan!" @StockSavvyShay: building the hardware into Toyota's factories "could significantly lower cost per mile as Waymo scales globally while opening door to bringing its tech into personally owned Toyotas."
Toyota paid $550M for Lyft's self-driving unit in 2021, and now it builds the same bZ4X robotaxi for Pony.ai in China and for Waymo in Japan. The biggest carmaker in the world has given up on its own driver and puts its factories (about 10M cars a year, when Waymo has about 4,000 on the road) behind whoever has the best one in each country.
Labs news & launches
1. [rumour] Optimus reservations in early December, says a summit organiser. Musk says no
Humanoids Summit organiser Modar Alaoui told the San Francisco Standard that Tesla plans to open consumer reservations around his 1-2 Dec summit, where it will demo Optimus V3, with most units going to companies at over $20,000 (@TheHumanoidHub). Asked if ordering was coming in December, Musk replied "No".
@tslaming laid out "a more realistic timeline": "B2B deliveries in late 2027 to 2028" and "Consumer sales in 2028+". @RealDanODowd: "Will it come with the VR headset and motion capture suit?"
Heat: medium · 758K views, 33 posts · quiet on Reddit
2. Agility Robotics starts trading on Nasdaq as AGLT on 5 Nov, with about $465M
Its SPAC merger with Churchill Capital XI should close on 4 Nov, pending a shareholder vote on 2 Nov. The $465M includes the SPAC's trust cash, a $201M private placement led by Foxconn and money already received (@IPONewsroom_, @humanoidsdaily). Churchill's stock fell about 18% because over $620M had been expected.
@mikekalilmfg: "Whether it's a good long-term investment depends on whether the funding is enough for Agility to get Digit 5 into working shape". @xmercury_one: "Every existing holder, NVIDIA, Amazon, SoftBank and Schaeffler included, rolls all its equity in under a 180 day lockup."
Heat: low · 29K views, 27 posts · quiet on Reddit
3. Lightberry's Lumi is the first humanoid cleared for US sale under the ban on foreign-made robots
After a Department of War national security review, Lumi got conditional approval and an exemption from the FCC's Covered List, which has blocked new foreign-made robots since July. The conditions: Lightberry replaces all sensors, compute and software with its own, and builds the body, head and hands in the US (@TheHumanoidHub). The 4'2" robot costs $39,990.
@ycombinator: "the first and currently only humanoid company in the US that can ship robots - while using an existing supply chain." @theobot__: "Sorry but the robot is literally a Unitree G1".
Heat: medium · 110K views, 19 posts · quiet on Reddit
4. Axonnic's sleeve reads forearm muscles to record how skilled hands work
The YC F26 startup's EMG sleeve tracks finger movement and estimates fingertip force while people work, synced with video and hand pose, as training data for dexterous robots. It keeps decoding the fingers when the camera loses sight of the hand, it says. Sleeves are shipping, with paid pilots under way.
No takes yet.
Heat: low · 8.1K views, 1 post · 60 votes on YC Launches
Key research
Ten frontier models drove a surgical robot in simulation, and only two ever finished a task
In micro1's new Dual-Arm Sim, across 800 trials on four tasks (from swapping surgical instruments to replacing a drug cartridge), Gemini 4 Argon finished nine and GPT-6 Astra two. Models said "task complete" 213 times and were wrong every time; in 94 the robot never touched the object (@TAAFTResearch). It matters because general-purpose models miss the grasp and then report success anyway, so a robot's sensors have to check every step.
UMI-Live: an open-source two-handed rig that goes from recording demos to a trained policy
Each handheld gripper carries its own camera, LiDAR, compute and battery, so nothing needs calibrating, and episodes become LeRobot datasets on Hugging Face while you record. VLAs fine-tuned mostly on its data succeeded in 10 of 10 trials (π0.5) and 8 of 10 (MolmoAct2) on a real bimanual task, the authors report. Hardware, code and about 40 hours of data are open. It matters because any lab can copy the full path from handheld demos to a trained policy, with nothing to calibrate in the room.
MolmoSpaces-CommonSense: no robot policy tops 30% when a task needs common sense
The benchmark hides the mug behind other objects, or asks for "the hot pan", to test commonsense reasoning in manipulation. No policy tested topped 30% on any task, and the best succeeded 9.3% of the time (@DJiafei). It matters because policies that look great on today's benchmarks fall apart here, so it gives labs a hard target they're nowhere near saturating.