Core Matter (@CoreMatter_) header
Core Matter (@CoreMatter_) profile picture

The Business of Robotics and Physical AI. Research & Podcast. https://xitter.gq/t.co/aanMm8guOj

Joined June 2026
Photo by Core Matter (@CoreMatter_)Photo by Core Matter (@CoreMatter_)Photo by Core Matter (@CoreMatter_)Photo by Core Matter (@CoreMatter_)Photo by Core Matter (@CoreMatter_)Photo by Core Matter (@CoreMatter_)Photo by Core Matter (@CoreMatter_)Photo by Core Matter (@CoreMatter_)Photo by Core Matter (@CoreMatter_)Photo by Core Matter (@CoreMatter_)
Core Matter studies the business of Physical AI. We publish research on robotics costs, supply chains and deployment, plus interviews with the people building the stack. Follow for research, charts and interview clips. Start here in the reply.
1
1
4
583
In one robot training test, 10,000 hours of glove recordings beat 2mn hours of web video. Beijing startup @liberai_robot ran the test for its Liber-0 model, released the week before CoRL 2026: โ€ข ๐—˜๐—ฎ๐—ฐ๐—ต ๐—ฑ๐—ฎ๐˜๐—ฎ ๐˜€๐—ผ๐˜‚๐—ฟ๐—ฐ๐—ฒ ๐—ด๐—ผ๐˜ ๐—ฎ ๐—ณ๐—ฎ๐—ถ๐—ฟ ๐˜๐—ฒ๐˜€๐˜. LiberAI pretrained separate models on 5 data sources, then gave each the same robot post-training and the same 76 tasks. โ€ข ๐—ช๐—ฒ๐—ฏ ๐˜ƒ๐—ถ๐—ฑ๐—ฒ๐—ผ ๐—ฎ๐˜ ๐Ÿฎ๐—บ๐—ป ๐—ต๐—ผ๐˜‚๐—ฟ๐˜€ ๐˜€๐—ฐ๐—ผ๐—ฟ๐—ฒ๐—ฑ ๐Ÿฒ๐Ÿฐ.๐Ÿฒ ๐—ฎ๐—ป๐—ฑ ๐Ÿฑ๐Ÿญ.๐Ÿฎ. The first score is pick-and-place and the second is harder manipulation, on a progress score from 0 to 100. โ€ข ๐—š๐—น๐—ผ๐˜ƒ๐—ฒ ๐—ฟ๐—ฒ๐—ฐ๐—ผ๐—ฟ๐—ฑ๐—ถ๐—ป๐—ด๐˜€ ๐—ฎ๐˜ ๐Ÿญ๐Ÿฌ,๐Ÿฌ๐Ÿฌ๐Ÿฌ ๐—ต๐—ผ๐˜‚๐—ฟ๐˜€ ๐˜€๐—ฐ๐—ผ๐—ฟ๐—ฒ๐—ฑ ๐Ÿฒ๐Ÿณ.๐Ÿต ๐—ฎ๐—ป๐—ฑ ๐Ÿฑ๐Ÿต.๐Ÿต. The glove measures the hand's motion alongside the video. โ€ข ๐—ง๐—ต๐—ฒ ๐—ด๐—ฎ๐—ฝ ๐˜„๐—ฎ๐˜€ ๐˜„๐—ถ๐—ฑ๐—ฒ๐˜€๐˜ ๐—ผ๐—ป ๐—ต๐—ฎ๐—ฟ๐—ฑ๐—ฒ๐—ฟ ๐—บ๐—ฎ๐—ป๐—ถ๐—ฝ๐˜‚๐—น๐—ฎ๐˜๐—ถ๐—ผ๐—ป. On tasks such as folding and opening, glove data kept improving as hours grew. โ€ข ๐—–๐—ผ๐—บ๐—ฏ๐—ถ๐—ป๐—ถ๐—ป๐—ด ๐—ฏ๐—ผ๐˜๐—ต ๐˜€๐—ฐ๐—ผ๐—ฟ๐—ฒ๐—ฑ ๐—ต๐—ถ๐—ด๐—ต๐—ฒ๐—ฟ ๐˜€๐˜๐—ถ๐—น๐—น. Liber-0's full mix of 2.2mn hours scored 83.6 and 73.8. โ€ข ๐—ก๐—ฉ๐—œ๐——๐—œ๐—”'๐˜€ ๐—š๐—ฅ๐Ÿฌ๐Ÿฌ๐—ง ๐Ÿฎ ๐˜๐—ฎ๐—ธ๐—ฒ๐˜€ ๐—ฎ ๐˜€๐—ถ๐—บ๐—ถ๐—น๐—ฎ๐—ฟ ๐—ฎ๐—ฝ๐—ฝ๐—ฟ๐—ผ๐—ฎ๐—ฐ๐—ต. It labels its human video with tracked 3D hand poses. For anyone buying robot training data, how the motion is recorded matters as much as how many hours it covers.
The robot AI stack is splitting in two. CoRL 2026 showed it clearly. (1) A planner decides what to do, and (2) a fast model, the actor, does the hands-on work. @OpenAI's GPT-6 Astra planned while @nvidia's GR00T 2 handled the manipulation, and the pair succeeded in 97 to 98 of 100 toy assembly trials. @GoogleDeepMind used the same split on Mercedes-Benz lines. 5 key takeaways from CoRL 2026: 1. The planner layer is getting cheaper. Luma AI's open 27bn-parameter model matched Astra at $0.29 per task, against $4.87. 2. Training data is getting cheaper. GR00T 2 trained on about 92% human video and about 5% robot teleoperation. 3. How the motion is recorded matters more than raw hours. In LiberAI's test, 10,000 hours of glove recordings beat 2mn hours of web video. 4. Deployment data compounds. Russ Tedrake called mid-training data "commoditized" and said deployment data is something "you have to earn." 5. Touch is the next frontier. @robbyant_brain open-sourced LingBot-Touch, a tactile model pretrained on nearly 5,000 hours of real tactile data. Full #CoRL2026 recap on @CoreMatter_: read.corematter.com/p/corl-2โ€ฆ
1
1
4
148
Is any robot training data still hard to find? Oliver Cameron (@olivercameron), Co-founder & CEO of @odysseyml: "Not anymore." โ€ข Odyssey stopped sending people out with camera backpacks โ€ข It buys from 10 to 12 data vendors โ€ข For an unusual request, 3 of them would start the next day โ€ข Next, Oliver expects world models to become training environments and data generators for other AI
INSIDE @odysseyml, the $1.45bn world model lab, w/ Co-founder & CEO Oliver Cameron (@olivercameron) Compute deals agreed on a handshake disappeared the next day. "Deals that we had handshaken were the next day disappearing." Odyssey raised a $310mn Series B in June, with @amazon and @AMD in the round. It trains across @nvidia, @awscloud Trainium and AMD chips. "I think there's equally a case to be made that it will actually get dramatically worse, that compute will become more scarce, will become more expensive, more constrained." We cover: โ€บ How @OpenAI and @AnthropicAI raising huge rounds squeezed compute for smaller labs โ€บ Why the Series B became a search for compute partners โ€บ How coding agents make models portable across chips โ€บ Robot data is abundant now. 3 of Odyssey's 10 to 12 vendors would start an unusual request the next day โ€บ World models as training environments and data generators โ€บ The Voyage lesson: solve the hardest version of the problem first TIMESTAMPS 00:00 What is Odyssey-3, a foundation world model? 02:05 Why task-specific robot data hits a ceiling 04:50 Driving in India: one researcher, results of a 500-person team 06:57 20 hours of blocky simulation vs 30 hours of real driving 09:21 Flexion, post-training and why robot companies stop building everything 11:45 Poke & Wiggle (@pokeandwiggle): real-robot evals as a test suite 15:15 World models vs language models: 14B parameters vs a 5T rumor 18:48 What it takes to scale world models 21:37 Egocentric video and the data mixture 23:33 Is any robot data still hard to find? 25:26 World models as the ultimate RL environment 27:30 When compute became existential 29:29 Fundraising as a compute partnership search 31:44 A hybrid fleet: NVIDIA, Trainium and AMD 33:13 Coding agents and CUDA lock-in 34:18 What is a robot company's moat? 36:17 A short memory for pain 37:55 The Voyage lesson: solve the hardest thing first
2
111
The missing piece in US reindustrialization: machines that build the machines
Who builds the machines that build the machines? I went down the rabbit hole of US machine tools: Imagine the Russian Doll of Physical AI bottlenecks: It's the machines that cut, grind and shape the parts inside aircraft, turbines and robots. โ€ข The US imported about half of the machine tools it used in 2025, $5.8bn out of $11.4bn. US output would need to grow 63% to 83% to replace imports with domestic machines. โ€ข Precision machine tools are a key area of import dependence. Imported gear grinders help make the small gears inside robot joints and aerospace gearboxes. โ€ข US orders for metalworking machinery rose to the highest value since 1998, up 45% in Jan to Sep 2026, reaching $5.76bn. The full piece maps what the US builds, what it imports and what expanding domestic production takes, with Section 232 as the backdrop. It also introduces our first @CoreMatter_ industry model, with US machine tool demand broken down by end market. read.corematter.com/p/sectioโ€ฆ
3
84
Robot training data is moving from robots to people to software. The cost of an hour of training data now depends on where it comes from: โ€ข ๐—ฅ๐—ผ๐—ฏ๐—ผ๐˜๐˜€ ๐—ฐ๐—ฎ๐—บ๐—ฒ ๐—ณ๐—ถ๐—ฟ๐˜€๐˜. In 2022, Google's RT-1 took 13 robots and 17 months to collect 130,000 demonstrations. โ€ข ๐—ง๐—ฒ๐—น๐—ฒ๐—ผ๐—ฝ๐—ฒ๐—ฟ๐—ฎ๐˜๐—ถ๐—ผ๐—ป ๐—ถ๐˜€ ๐—ป๐—ผ๐˜„ ๐˜€๐—ผ๐—น๐—ฑ ๐—ฏ๐˜† ๐˜๐—ต๐—ฒ ๐—ต๐—ผ๐˜‚๐—ฟ. Chinese data factories sell robot teleoperation data at Rmb500 to Rmb1,000 ($70 to $141) per hour. โ€ข ๐—ฃ๐—ฒ๐—ผ๐—ฝ๐—น๐—ฒ ๐˜„๐—ถ๐˜๐—ต ๐—ต๐—ฎ๐—ป๐—ฑ๐—ต๐—ฒ๐—น๐—ฑ ๐—ด๐—ฟ๐—ถ๐—ฝ๐—ฝ๐—ฒ๐—ฟ๐˜€ ๐—ฎ๐—ฟ๐—ฒ ๐—ณ๐—ฎ๐˜€๐˜๐—ฒ๐—ฟ. Japan's AIRoA had 179 people record 8,434 hours in a little over 2 months, with no robots involved. โ€ข ๐—›๐˜‚๐—บ๐—ฎ๐—ป ๐˜ƒ๐—ถ๐—ฑ๐—ฒ๐—ผ ๐—ถ๐˜€ ๐—ฐ๐—ต๐—ฒ๐—ฎ๐—ฝ๐—ฒ๐—ฟ ๐˜€๐˜๐—ถ๐—น๐—น. Mercor pays businesses $12 to $18 per approved hour of workplace footage, and @NVIDIARobotics' GR00T 2 trained on 402,730 hours of human video. โ€ข ๐—ฆ๐—ถ๐—บ๐˜‚๐—น๐—ฎ๐˜๐—ถ๐—ผ๐—ป ๐—ฐ๐—ผ๐˜€๐˜๐˜€ ๐˜๐—ต๐—ฒ ๐—น๐—ฒ๐—ฎ๐˜€๐˜. The @allen_ai MolmoB0T team estimates $2 to $5 per hour of simulated robot data. Robot teleoperation was about 5% of GR00T 2's pretraining hours.
The robot AI stack is splitting in two. CoRL 2026 showed it clearly. (1) A planner decides what to do, and (2) a fast model, the actor, does the hands-on work. @OpenAI's GPT-6 Astra planned while @nvidia's GR00T 2 handled the manipulation, and the pair succeeded in 97 to 98 of 100 toy assembly trials. @GoogleDeepMind used the same split on Mercedes-Benz lines. 5 key takeaways from CoRL 2026: 1. The planner layer is getting cheaper. Luma AI's open 27bn-parameter model matched Astra at $0.29 per task, against $4.87. 2. Training data is getting cheaper. GR00T 2 trained on about 92% human video and about 5% robot teleoperation. 3. How the motion is recorded matters more than raw hours. In LiberAI's test, 10,000 hours of glove recordings beat 2mn hours of web video. 4. Deployment data compounds. Russ Tedrake called mid-training data "commoditized" and said deployment data is something "you have to earn." 5. Touch is the next frontier. @robbyant_brain open-sourced LingBot-Touch, a tactile model pretrained on nearly 5,000 hours of real tactile data. Full #CoRL2026 recap on @CoreMatter_: read.corematter.com/p/corl-2โ€ฆ
1
66
Robots are being built in 3 layers, and each layer works on a different clock. Tsinghua professor Hang Zhao, co-founder of Galaxea, described this framework in his CoRL 2026 keynote, and @NVIDIARobotics built GR00T 2 the same way. โ€ข ๐—ฆ๐˜†๐˜€๐˜๐—ฒ๐—บ ๐Ÿฎ ๐—ฝ๐—น๐—ฎ๐—ป๐˜€ ๐—ผ๐˜ƒ๐—ฒ๐—ฟ ๐˜€๐—ฒ๐—ฐ๐—ผ๐—ป๐—ฑ๐˜€ ๐˜๐—ผ ๐—บ๐—ถ๐—ป๐˜‚๐˜๐—ฒ๐˜€. @OpenAI's GPT-6 Astra and @GoogleDeepMind's Gemini Robotics-ER work at this layer. โ€ข ๐—ฆ๐˜†๐˜€๐˜๐—ฒ๐—บ ๐Ÿญ ๐—ต๐—ฎ๐—ป๐—ฑ๐—น๐—ฒ๐˜€ ๐—ฑ๐—ฒ๐˜…๐˜๐—ฒ๐—ฟ๐—ถ๐˜๐˜†, ๐˜๐—ฟ๐—ฎ๐—ฐ๐—ธ๐—ถ๐—ป๐—ด ๐˜๐—ต๐—ฒ ๐—ต๐—ฎ๐—ป๐—ฑ ๐—ฎ๐—ป๐—ฑ ๐˜๐—ต๐—ฒ ๐—ผ๐—ฏ๐—ท๐—ฒ๐—ฐ๐˜ ๐˜๐—ฒ๐—ป๐˜€ ๐—ผ๐—ณ ๐˜๐—ถ๐—บ๐—ฒ๐˜€ ๐—ฝ๐—ฒ๐—ฟ ๐˜€๐—ฒ๐—ฐ๐—ผ๐—ป๐—ฑ. VLAs and world action models such as GR00T 2 compete here. โ€ข ๐—ฆ๐˜†๐˜€๐˜๐—ฒ๐—บ ๐Ÿฌ ๐—ฟ๐—ฒ๐—ฎ๐—ฐ๐˜๐˜€ ๐˜๐—ผ ๐—ฐ๐—ผ๐—ป๐˜๐—ฎ๐—ฐ๐˜ ๐—ฎ๐—ป๐—ฑ ๐—ณ๐—ผ๐—ฟ๐—ฐ๐—ฒ ๐—ต๐˜‚๐—ป๐—ฑ๐—ฟ๐—ฒ๐—ฑ๐˜€ ๐—ผ๐—ณ ๐˜๐—ถ๐—บ๐—ฒ๐˜€ ๐—ฝ๐—ฒ๐—ฟ ๐˜€๐—ฒ๐—ฐ๐—ผ๐—ป๐—ฑ. NVIDIA's SONIC whole-body controller sits here. โ€ข ๐—ก๐—ฉ๐—œ๐——๐—œ๐—”'๐˜€ ๐˜€๐˜๐—ฎ๐—ฐ๐—ธ ๐˜€๐—ต๐—ผ๐˜„๐˜€ ๐˜๐—ต๐—ฒ ๐—ฐ๐—น๐—ผ๐—ฐ๐—ธ๐˜€. GR00T 2 Turbo plans 1.6 seconds of motion about once a second on a Jetson AGX Thor, and SONIC sets joint targets 50 times a second. โ€ข ๐—ฆ๐—ฝ๐—ฒ๐—ฒ๐—ฑ ๐—ฐ๐—ต๐—ฎ๐—ป๐—ด๐—ฒ๐˜€ ๐˜๐—ต๐—ฒ ๐—ฟ๐—ฒ๐˜€๐˜‚๐—น๐˜๐˜€. In Zhao's simulated kitchen test, 4 approaches completed over 60% of tasks when thinking time was free. Once thinking time counted toward a 50-second deadline, a policy calling a frontier model completed 15%. Roboticist Benjie Holson called System 1 "the big challenge for robotics in the next 6 months."
The robot AI stack is splitting in two. CoRL 2026 showed it clearly. (1) A planner decides what to do, and (2) a fast model, the actor, does the hands-on work. @OpenAI's GPT-6 Astra planned while @nvidia's GR00T 2 handled the manipulation, and the pair succeeded in 97 to 98 of 100 toy assembly trials. @GoogleDeepMind used the same split on Mercedes-Benz lines. 5 key takeaways from CoRL 2026: 1. The planner layer is getting cheaper. Luma AI's open 27bn-parameter model matched Astra at $0.29 per task, against $4.87. 2. Training data is getting cheaper. GR00T 2 trained on about 92% human video and about 5% robot teleoperation. 3. How the motion is recorded matters more than raw hours. In LiberAI's test, 10,000 hours of glove recordings beat 2mn hours of web video. 4. Deployment data compounds. Russ Tedrake called mid-training data "commoditized" and said deployment data is something "you have to earn." 5. Touch is the next frontier. @robbyant_brain open-sourced LingBot-Touch, a tactile model pretrained on nearly 5,000 hours of real tactile data. Full #CoRL2026 recap on @CoreMatter_: read.corematter.com/p/corl-2โ€ฆ
1
1
4
276
Gift guide curated for the robotics and physical AI community. Let us know which ones are your favorites!
The holidays are around the corner. Inspired by @lennysan, I'm releasing the first Core Matter Holiday Gift Guide! The list is tailored for the hardware builders, the robot deployers, the Physical AI-curious software engineers, parents with kids who love robots, and even your suppliers in Shenzhen. It ranges from $20,000 to $14. From practical, professional tools, to wellness, delightful things that bring a smile to my face. Thanks to everyone who replied to the polls, including @alexembodied, @darthur, @benbeerengineer, @dcsan, @liminalsunset_, @TouficJrab, @boxcardavid, @_baldtires, @MikeAHFerguson, @Joyingeth, @RohitSatyaV, @reticente07, @LiamTal and @lasermatts. No affiliate links, so I'm not benefiting from recommending these products. Full guide here: read.corematter.com/p/robotiโ€ฆ
61
Core Matter retweeted
The holidays are around the corner. Inspired by @lennysan, I'm releasing the first Core Matter Holiday Gift Guide! The list is tailored for the hardware builders, the robot deployers, the Physical AI-curious software engineers, parents with kids who love robots, and even your suppliers in Shenzhen. It ranges from $20,000 to $14. From practical, professional tools, to wellness, delightful things that bring a smile to my face. Thanks to everyone who replied to the polls, including @alexembodied, @darthur, @benbeerengineer, @dcsan, @liminalsunset_, @TouficJrab, @boxcardavid, @_baldtires, @MikeAHFerguson, @Joyingeth, @RohitSatyaV, @reticente07, @LiamTal and @lasermatts. No affiliate links, so I'm not benefiting from recommending these products. Full guide here: read.corematter.com/p/robotiโ€ฆ
2
2
1
14
1,057
When you watch @Tesla Optimus handle an object, look at how the thumb meets the fingers and what keeps the grasp stable. Prensilia (@prensilia) Managing Director Francesco Clemente offers a crash course in how hand designers work backward from everyday tasks: - Start with the grasp. A cylindrical grip, a pinch grip and a lateral grip use the hand differently. - Study which grasps people use. Francesco describes looking at grasp taxonomies to understand the hand shapes needed in daily activities. - Design around the task. The useful set of movements depends on the work the hand needs to perform. - Consider manufacturing complexity. A hand with many motors can perform many movements, but Francesco explains why scaling that design becomes difficult. - Choose coverage deliberately. Prensilia studies how a simpler mechanism can cover commonly used grasps while keeping cost and complexity manageable. Watch Prensiliaโ€™s managing director explain how those choices shape a robot hand.
111
Following @1x_techโ€™s NEO and the push toward robots in the home? In San Francisco, people can pay $60 for two @taurobots robots to clean their home for an hour. During a @CoreMatter_ visit: - They completed about 80% of one floor. Tasks included tidying, wiping surfaces, vacuuming, mopping and cleaning the toilet. - Each robot had a remote human operator. Staff also handled delivery and pickup. - Delicate tasks took more time. Moving objects to clean underneath them and handling soft items were difficult. - Batteries needed swapping after about an hour. A longer booking would need a plan for that turnaround. - The visit also supplied training data. Data collection was accepted as part of the booking to help train the robots. Would you let robots clean your home?
2
117
Building hardware at @Apple and ordering your first small production run are different jobs. Seeed Studio (@seeedstudio) founder Eric Pan (@256ericpan) describes what changes when experienced corporate hardware teams become startup founders: - Order volume changes the relationship. A startup may bring demanding specifications while placing a small order with a small supplier. - Corporate methods need adapting. Eric describes founders with @Apple or @Microsoft experience discovering that the approach from their former employer does not fit the new situation. - Collaboration becomes central. He encourages a more agile, relationship-based way of working with manufacturers, with room to discuss constraints and adjust together. - Requirements need context. The question is what this supplier can deliver for this product at this volume, with the resources available. For a founder making that transition, which part of your manufacturing approach would you change first? Watch Seeed Studioโ€™s founder explain the shift.
1
85
A farm machine carrying 18 desktop computers caught Bonsai Robotics (@bonsairobotics) CEO Tyler Niday (@niday_tyler)โ€™s attention. The machine came from Blue River Technology, which later became part of John Deere (@JohnDeere). Reflecting on his early work at Blue River, he recalls: - A specific job: the tractor thinned lettuce while moving through the field, with the computers mounted on its back. - A later computing platform: the team used @nvidiaโ€™s TX1 with YOLO-based deep learning as its work evolved. - A practical perception problem: distinguishing green weeds from green crops so the machine could tell them apart. Watch Bonsai Roboticsโ€™ cofounder recount the work.
1
1
77
A humanoidโ€™s knee and wrist need different actuator properties. They perform different jobs inside the same robot. Our actuator graphic maps typical choices across joint groups. - Legs manage changing loads and repeated ground contact. - Wrists need precision within limited packaging space. - Weight near the armโ€™s end adds another design constraint. Start with the jointโ€™s job, then examine its actuator. The graphic shows typical trade-offs across designs.
2
2
91
Manufacturing speed starts before a factory builds the product. Eric Pan, Seeed Studioโ€™s founder, describes the surrounding network. Hardware teams face trade-offs between quality, cost and lead time. They also need access to people and specialized capabilities. - Engineers bring the skills needed for the project. - Suppliers provide the components and capabilities it requires. - Their responsiveness helps the team move as requirements change. Pan describes that combination in Shenzhenโ€™s manufacturing network. Finding the right capabilities is part of getting hardware built.
2
2
1
4
2,698
A robot hand contains several businesses inside one product. Complete-hand manufacturers bring together components from several technical disciplines. Our July supply-chain graphic separates those layers. - Actuation supplies motion through motors and related components. - Transmission carries that motion through the handโ€™s mechanisms. - Sensors provide information about movement and contact. - Control electronics use that feedback to coordinate movement. The map highlights component opportunities alongside complete-hand companies. Its company examples reflect the July primer.
1
57
A transparent cup can confuse a robotโ€™s depth camera. Its depth map can contain very little usable information. Shan Luo explains how touch helps fill that gap. He studies robotic perception at Kingโ€™s College London. - Vision identifies sparse cues, such as the cupโ€™s rim. - Those cues guide where the robot makes contact. - Touch checks the edges and helps estimate the cupโ€™s position. In these experiments, vision guides touch before the grasp. Contact gives the robot information that vision struggles to provide.
1
25