Coverage on the next frontier of AI. Robotics and Physical AI events, media & interviews. The waitlist for AUTONOMOUS 2027 is now live.
San Francisco, CA
Joined April 2026
- Tweets1.1K
- Following114
- Followers368
- Likes2.3K
Egocentric video data is "booming"...but for how long?
The last piece yet to be solved is tactile data.
@JayLiStanford's vision is a future where the majority of mass data is still purely vision, but the higher quality, lower quantity data is only possible with touch included.
Watch the full podcast below π
What happens if the brains outperform the hands?
Hardware falls behind, failing to keep pace.
Proception builds to close this gap. Their ProHand, a tendon-driven hand with 22 degrees of freedom in a human form factor, and ProGlove, a sensor glove that captures the touch data egocentric video cannot, is how they do this. In this episode, we talk to Jay Li, Founder and CEO, about why he thinks hands are now the bottleneck for the whole robotics stack.
This year, they're ramping production from single digits a month to hundreds. We cover why tendons and five fingers, how glove data compares with teleoperation, a first model trained with zero teleoperation and zero robot data, and how Jay expects the humanoid market to split.
@JayLiStanford @proceptionAI
Full discussion π
01:44 The 2027 humanoid prototype and the hands business
06:51 Scaling to 100s of hands a month
11:41 Why tendons
13:22 Why five fingers, not three or four
16:18 Where tactile data fits alongside egocentric video
19:58 Why the glove carries 96 sensors
22:38 Closing the embodiment gap
25:12 A tactile model trained with zero teleoperation
27:15 Glove data versus teleoperation on cost and scale
31:18 Why hands are now the bottleneck
37:26 Will humanoids consolidate like LLMs
42:00 What has kept the team going
Google DeepMind, NVIDIA & a $14B startup called Skild AI are all building the same thing: one AI brain that can run any robot.
We mapped 160+ companies building that brain & everything it needs to work.
β³ Brain builders like @GoogleDeepMind & @NVIDIARobotics
β³ Robot makers training their own, like @Figure_robot, @1x_tech & @UnitreeRobotics
β³ World models, simulation & training data to teach them
β³ The chips & software that run them on the robotTwo of the biggest bets came out of university labs. @chelseabfinn & @svlevine started @physical_int, & @deepakpathak started @SkildAI.
Who are we missing?
PS - Our 2027 waitlist is live now: autonomousfuture.co
2
133
Google DeepMind's new robot brain is already sorting car parts on Mercedes-Benz production lines, with a 96% success rate.
Gemini Robotics 2.5 is its most capable robot model yet, tested on 200 tasks, from tying a bow on a gift box to pipetting liquid.
β³ 30% to 100% higher success than its previous models, while finishing tasks 20% to 30% faster
β³ Running on @Apptronik's Apollo 2 humanoid, it sorted parts into kitting trays at Mercedes-Benz plants in Hungary & Germany
Built by @GoogleDeepMind, with a training boost from Gemini 4 Argon, which turns a single camera view into thousands of simulated practice runs.
π₯ @GoogleDeepMind
@GoogleDeepMind @Apptronik βThis is what we built Apollo for: useful work alongside people in real industrial settings.β Jeff Cardenas, Apptronik CEO
Read more here: deepmind.google/blog/gemini-β¦
1
33
Amazon's Ring just launched robots that patrol for you: a $249.99 drone that flies around your home & a $999.99 rover for businesses.
Fixed cameras only see one spot. Ring Robotics sends the camera to whatever set off the alarm.
β³ Ring Drone Cam flies preset routes indoors, only records once it takes off, & ships Jan 27, 2027
β³ Ring Rover Cam is built on Amazon's Astro tech, records 4K on the move with a periscope camera, & ships Dec 16, 2026
Both run on AI Pilot, which maps the space, suggests patrol points & steers around obstacles. Built by @ring, led by founder & chief inventor @JamieSiminoff, under Amazon devices chief @panos_panay.
π₯ @panos_panay
Self-reply
@ring @JamieSiminoff @panos_panay βIntroducing Ring Robotics, a new category of autonomous security launching with two products: Ring Drone Cam and Ring Rover Cam.β
Read more here: aboutamazon.com/news/devicesβ¦
3
27
The AI welding robots working for America's largest shipbuilder just raised $300M at a $2.45B valuation.
Path Robotics started with welding, one of the hardest factory jobs to automate. Now it wants every task before & after the weld too.
β³ Annual recurring revenue grew 17x in the last 18 months
β³ Customers include HII, Lockheed Martin & Saronic, plus mid-sized shops like TYCROP & Millerbernd
The Series E was led by @Georgian_io, with Tiger Global, Drive Capital, Addition, D. E. Shaw Ventures, NAVER Ventures & Yamaha Motor Ventures. @PathRobotics is led by co-founder & CEO Andy Lonsberry.
π₯ @PathRobotics
@PathRobotics @Georgian_io βWe started with the hardest job in the factory: welding.β
Read more here: path-robotics.com/news/path-β¦
2
24
Ant Group's robotics arm just open-sourced a model that lets robot hands feel, trained on nearly 5,000 hours of real touch data.
Robot hands mostly work by sight. With LingBot-Touch, a robot unscrewing a bottle cap went from 56% success to 90%.
β³ One model reads 3 kinds of touch sensors: magnetic, visual & piezoresistive
β³ Tested on 26 real tasks across 2 hands, Sharpa Wave & XHand
β³ Lifted average success by 14 to 26 points across 4 action models, including Physical Intelligence's Ο0.5
β³ The tiny version has about 5.3M parameters & adds only 0.13 to 0.16% to a robot's model size
β³ Code & technical report are free on GitHub & Hugging Face
Built by @robbyant_brain, Ant Group's embodied AI company.
"With LingBot-Touch, sensor and dexterous-hand manufacturers can select hardware based on cost and task requirements, while robotics teams can retain their own action policies."
π₯ @robbyant_brain
1
6
132
@robbyant_brain "What changes when robots can feelβnot just see? π€β"
Read more here: businesswire.com/news/home/2β¦
2
15
The ENTIRE humanoid industry might only have 1 or 2 vertically integrated companies.
Everyone else will be an 'ecosystem player'.
@JayLiStanford joined us for this week's AUTONOMOUS.
He believes we won't see consolidation among 5 large players because robotics will never scale the way LLMs did. The sheer size of the robotics market will mean that whilst few vertical humanoid companies exist, the market will instead be made up of different companies which use third parties for hardware.
We will see the industry shape itself to become more like PCs or autonomous vehicles.
Do you agree with this?
What happens if the brains outperform the hands?
Hardware falls behind, failing to keep pace.
Proception builds to close this gap. Their ProHand, a tendon-driven hand with 22 degrees of freedom in a human form factor, and ProGlove, a sensor glove that captures the touch data egocentric video cannot, is how they do this. In this episode, we talk to Jay Li, Founder and CEO, about why he thinks hands are now the bottleneck for the whole robotics stack.
This year, they're ramping production from single digits a month to hundreds. We cover why tendons and five fingers, how glove data compares with teleoperation, a first model trained with zero teleoperation and zero robot data, and how Jay expects the humanoid market to split.
@JayLiStanford @proceptionAI
Full discussion π
01:44 The 2027 humanoid prototype and the hands business
06:51 Scaling to 100s of hands a month
11:41 Why tendons
13:22 Why five fingers, not three or four
16:18 Where tactile data fits alongside egocentric video
19:58 Why the glove carries 96 sensors
22:38 Closing the embodiment gap
25:12 A tactile model trained with zero teleoperation
27:15 Glove data versus teleoperation on cost and scale
31:18 Why hands are now the bottleneck
37:26 Will humanoids consolidate like LLMs
42:00 What has kept the team going
@Nominal_io, just raised $240M led by @Lightspeedvp & co-led by @Insightpartnersat at a $3B valuation.
In March, it raised at $1B. 8 months later, its valuation has tripled.
β³ Hardware teams use it to pull test data from aircraft, robots, spacecraft & energy systems into one place
β³ Customers include Anduril, Scout AI, REGENT & Antares, plus 5 of the 7 biggest defence contractors
β³ Its first AI agent, Nominal Analyst, is built for hardware engineers
β³ $395M raised across its last 3 rounds
The software is used by Anduril to test its machines.
Also joining the round are Sequoia, Founders Fund, General Catalyst & Lockheed Martin Ventures.
π₯ @Nominal_io
Nominal is led by co-founder & CEO Cameron McCord.
@Nominal_io @Lightspeedvp @Insightpartners βEngineers today are building some of the most ambitious machines humanity has ever seen. They should have AI thatβs purpose-built for them.β
Read more here: nominal.io/blog/series-c
2
22
What happens if the brains outperform the hands?
Hardware falls behind, failing to keep pace.
Proception builds to close this gap. Their ProHand, a tendon-driven hand with 22 degrees of freedom in a human form factor, and ProGlove, a sensor glove that captures the touch data egocentric video cannot, is how they do this. In this episode, we talk to Jay Li, Founder and CEO, about why he thinks hands are now the bottleneck for the whole robotics stack.
This year, they're ramping production from single digits a month to hundreds. We cover why tendons and five fingers, how glove data compares with teleoperation, a first model trained with zero teleoperation and zero robot data, and how Jay expects the humanoid market to split.
@JayLiStanford @proceptionAI
Full discussion π
01:44 The 2027 humanoid prototype and the hands business
06:51 Scaling to 100s of hands a month
11:41 Why tendons
13:22 Why five fingers, not three or four
16:18 Where tactile data fits alongside egocentric video
19:58 Why the glove carries 96 sensors
22:38 Closing the embodiment gap
25:12 A tactile model trained with zero teleoperation
27:15 Glove data versus teleoperation on cost and scale
31:18 Why hands are now the bottleneck
37:26 Will humanoids consolidate like LLMs
42:00 What has kept the team going
NVIDIA is planning to invest in RLWRLD, a startup whose robot model claims to beat NVIDIAβs own GR00T.
RLWRLD builds AI for five-finger robot hands, the kind that thread nuts onto bolts & pick parts off moving conveyors.
β³ Its RLDX-1 model beats GR00T & Physical Intelligenceβs Ο0 on 8 public benchmarks, per RLWRLD
β³ It trains on force, touch & the moment of contact, not vision alone
β³ $41M raised across 2 seed rounds, backed by CJ Logistics, Lotte & Kakao
β³ 7 papers at CoRL 2026 in Austin this week
@RLWRLD_ai is based in SF, Seoul & Tokyo and led by founder & CEO Junghee Ryu. @nvidia announced the planned investment today at its AI Startup Partnership Forum in Seoul.
βRobotics intelligence canβt be built in isolation, it has to be trained where work actually happens.β
π₯ @RLWRLD_ai
@RLWRLD_ai @nvidia βToday at AI Day Seoul, NVIDIA announced a planned investment in RLWRLD. We intend to put it toward advancing our robotics foundation models and scaling industrial deployments.β
Read more here: blogs.nvidia.com/blog/koreanβ¦
3
44
AUTONOMOUS
β
β£ π Hardware
β β£ π Humanoids
β β£ π Actuators & Motors
β β£ π Dexterous Hands
β β£ π Tactile Sensing
β β£ π Batteries & Power
β β£ π Drones & AVs
β β π Factory Robots
β
β£ π Software
β β£ π Physical AI
β β£ π World Models
β β£ π Sim2Real
β β£ π Teleoperation
β β£ π Robot Data
β β£ π VLAs & Foundation Models
β β π Fleet Deployment
β
β£ π Previous Speakers
β β£ π Rhoda AI
β β£ π Dusty Robotics
β β£ π Built Robotics
β β£ π FORT Robotics
β β£ π Carbon Robotics
β β£ π Torc Robotics
β β£ π Kodiak Robotics
β β£ π Anyscale
β β π + 45 more
β
β£ π Mission
β β£ π Build the home for robotics & physical AI
β β£ π Bring the embodied AI industry together
β β π Showcase the next frontier of AI
β
β π AUTONOMOUS 2027...loading
1
4
215
Point72 & a16z both backed robotics startups this week. $84M+ went into 7 pre-seed & seed rounds.
Corridian alone took 45% of it, & 3 of the 7 rounds went to factory & industrial-site robots.
Seed:
β³ @corridian_inc: $38M for autonomous underwater vehicles that carry shipping containers, led by @p72vc
β³ @PhysiclAI: $10M for synthetic data that trains robots on rare failure cases
β³ Gallium Defense: β¬7.1M (~$8.2M) for radar that spots drones
β³ @armatrix_IN: βΉ67.28 Cr (~$7.1M) for snake-like robot arms that reach inside tanks through manholes
β³ AFORMIC: $6.95M to take factory robots & autonomous forklifts into AI data centres
Pre-seed:
β³ @bySyncere: $11M for LUME, a robot lamp for the home, backed by @a16z
β³ @StriveRobotics: $3M+ for robots built for factories where production runs change often
Which round would you back?
Nio Robotics just unveiled Aru, a robot with no central body. Just arms.
Today crews grit blast steel inside tanks & pipe runs. Aru strips it with a laser instead: no grit, no blasting, no contaminated waste.
β³ Rover mode grips vertical & overhead steel with magnetic feet
β³ Snake mode slides through 450 mm hatches & 40 cm pipes with 90Β° bends
β³ Quadruped & tripod modes climb stairs & ladders, then plant for 15 kg of tool reach
β³ Laser decoating & airless painting are live, welding is in development
Built in Toulouse by @nio_robotics, led by founder & CEO @TitouanLM.
π₯ @nio_robotics
1
3
126
@nio_robotics @TitouanLM βWe needed to build something that could go everywhere...β
Aru started as a garage prototype 5 years ago & made its first public outing at CES Las Vegas.
Read more here: nio-robotics.com
2
48