Pinned Tweet
Replying to @Robot_AIsignals
@Robot_AIsignals @VisiveAI @Sebagyeong @AlexIljaszewicz quick invite: The Figure User Group is a new chat for people who plan to bring Figure home. Ideas, questions and feedback all welcome: teapawt.org/i/chat/group_join/g211…
3
272
Building a community of future Optimus owners to share what we learn and speak with one voice to Tesla and SpaceX. @OrlandoLorenzo @therustoracle @Riverwindview @Gfilche, care to join The Optimus Users Group? teapawt.org/i/chat/group_join/g211…
100
Building a community of future Figure owners to share what we learn and speak with one voice to Figure AI. @Brownstone_X @BasicProtein26 @entrelligence_ @21stRobotics, care to join The Figure User Group? teapawt.org/i/chat/group_join/g211…
14
Overlooked Memo post of the week: @nikhilkr's neat primer: two arms on a wheeled base reaching floor to 7 ft, an ACT-2 model trained on ~10M episodes from 500+ homes, and a soft, safety-first build.
Join The Memo Users Group: teapawt.org/i/chat/group_join/g211…
teapawt.org/nikhilkr/status/211077…
Humanoid of the day: Sunday Robotics Memo
Specifications:
> Two arms on a wheeled base with a telescoping column
> Reaches from the floor up to 7 ft
> Soft and silicone-like covering and a stable base
> Runs on Sunday’s own ACT-2 model
Capabilities:
> Can do most household chores like loading a dishwasher, folding laundry, tidying and even making an espresso
> Trained on about 10 million episodes from 500+ homes and now without per-home training, it can work with great accuracy.
Deployments:
> Founding Family Beta with 50 households and individually numbered robots is targeted for 2026
Developer friendliness:
> Earlier research like ALOHA, ACT, Diffusion Policy etc. is open but Memo is a consumer product
The safety design is the most impressive part of this humanoid.
Wheels, soft body, these are the kind of things that makes it safe in being near kids too.
Excited to see real deployments of this one across homes soon @sundayrobotics 🚀
84
Overlooked Figure post of the week: @_TheMysterio on new Figure 03s walking themselves out of BotQ, with the scale behind it: a first line built for up to 12,000 humanoids a year.
Join The Figure User Group: teapawt.org/i/chat/group_join/g211…
teapawt.org/_TheMysterio/status/21…
Figure's newest robots don't get shipped out. They walk out.
Fresh Figure 03 humanoids leaving the BotQ factory and crossing campus on foot.
BotQ's first line was built for up to 12,000 humanoids a year.
The 1,000th Figure 03 came off it in July.
teapawt.org/adcock_brett/status/21…
76
Overlooked NEO post of the week: @StartupContext's balanced thread on 1X. Hayward built for ~10k NEOs a year, a 50k goal for 2027, Expert Mode privacy, and why capacity isn't deliveries.
Join The NEO Users Group: teapawt.org/i/chat/group_join/g200…
teapawt.org/StartupContext/status/…
Replying to @1X_Technologies
@1X_Technologies wants you to pay $20,000 for a robot that does your household chores.
But when the Wall Street Journal tested NEO, a human wearing a VR headset controlled its complex tasks.
Now the OpenAI-backed startup is reportedly discussing a $6 BILLION valuation, with no independently confirmed deliveries to paying households.
Here's the story. 🧵
2
1
68
Overlooked Optimus post of the week: @Ultima_Lumina breaks down how Optimus dances. It's one learned whole-body policy trained on thousands of simulated bots, then moved to hardware zero-shot.
Join The Optimus Users Group: teapawt.org/i/chat/group_join/g211…
teapawt.org/Ultima_Lumina/status/2…
How Optimus Actually Dances (and Hands Out Candy) 🎃
Optimus does not run on a hardcoded joint by joint routine.
The dancing you see is in the real demos are a learned whole body policy: a single neural network outputting continuous motor commands to every actuator in real time based on its own sensors and high level goals.
The engine behind this is simulation based reinforcement learning with zero shot sim to real transfer.
Tesla runs thousands of virtual Optimus instances inside a physics simulation simultaneously.
Each virtual bot tries a movement, receives a reward signal for staying balanced, hitting the beat, using smooth torque, and keeping its feet planted, and the policy updates.
By applying domain randomization (constantly altering virtual friction, mass, actuator latency, and lighting), the network learns true dynamic stability rather than memorizing a single rigid simulation.
Once solid, those exact neural weights drop straight onto physical hardware without extra tuning. Engineers confirmed those EDM sequences were trained completely in sim and transferred zero shot.
That same policy solves the classic bipedal balance problem. Through joint encoders, IMUs, and visual feedback, the robot constantly calculates its center of mass, foot contact forces, and angular momentum.
Micro adjustments run through the ankles, hips, and torso at high frequency so it stays upright while arms and legs move expressively.
The shift from safety tethers in early demos to untethered performance comes down to improved simulation fidelity and expanded training data, not manual scripting.
Hardware makes the motion possible.
High torque actuators drive the legs and torso, lower inertia units power the arms, and current hand designs shift toward high degree of freedom, tendon driven fingers.
That allows the exact same platform that holds a rhythm to execute delicate tasks like gripping candy without crushing it.
This feeds into the same neural family powering Tesla vehicle autonomy, with the upcoming AI5 chip providing the compute headroom required for these high degree of freedom policies.
To scale task mastery, Tesla is developing the Optimus Academy model: deploying tens of thousands of production units into continuous self play to generate real world training data. Elon Musk compares the process to raising toddlers where robots practice, fail, and adapt and learn through play.
Dancing is simply a visible milestone of a stack designed to walk, navigate, and handle objects autonomously.
When you watch Optimus dance at an event and hand out candy, you are observing a whole body network using its navigation, vision, and dexterity stacks simultaneously.
Hands remain the primary hardware bottleneck, and full long horizon autonomy is still developing, but the performance is no puppet show. It is the real time execution of a general learning loop.
Happy Halloween! Opti is just better at staying upright while the rest of us trip over the porch steps 🎃🍬
#Halloween #Optimus #Tesla
#Dance #Opti #OptimusAcademy
*This video is ai and made by me*
1
2
199
This community is dedicated to owners, early adopters, and enthusiasts of @sundayrobotics Memo – the groundbreaking home robot from Sunday.ai. Here, we come together to:
• Share real-world experiences with Memo in everyday home life
• Exchange tips, tricks, and creative ways to teach Memo new chores and tasks
• Discuss software updates, AI improvements, and hardware features
• Troubleshoot issues, offer support, and collaborate on best practices
• Celebrate milestones, post photos/videos of Sunday Memo in action, and speculate on the future of household robotics
• Foster a positive, respectful environment focused on safety, privacy, and the exciting potential of safe robots in our homes.
Whether you’re awaiting your delivery, already a Beta Tester, or just passionate about Sunday.ai’s vision for abundant physical labor through androids – this is your hub. Join us in shaping how Sunday Robotics transforms daily life!
Calling Figure watchers @guilleflorvs @WallStreetB0Ts @paddleurway @NewsTongueX! The Figure User Group is up and running as a home for people planning to own one. Swap news, use cases and questions with us: teapawt.org/i/chat/group_join/g211…
2
30
Building a community of future Figure owners to share what we learn and speak with one voice to Figure AI. @ctorobotics @AtlasShrug1 @EvanKirstel @FuturLucide, care to join The Figure User Group? teapawt.org/i/chat/group_join/g211…
12
If you're counting down to your own Optimus, The Optimus Users Group is for you. @FutureThesisX @TheHumanoidHub @Ledge_Andary @QAS_LLC, we'd love your perspective in there: teapawt.org/i/chat/group_join/g211…
Building a community of future Figure owners to share what we learn and speak with one voice to Figure AI. @wjkerley96 @yokoshaan @murzik52 @bot1kai, care to join The Figure User Group? teapawt.org/i/chat/group_join/g211…
16
If you're counting down to your own Figure, The Figure User Group is for you. @openkachi @seti_park @vanvinh2788 @leszoonex, we'd love your perspective in there: teapawt.org/i/chat/group_join/g211…
1
70
Overlooked Figure post of the week: @robixone looks past the headline on Figure 03's 9-day, 348-mile treadmill run and asks a fair question: does endurance or one day of chores convince you more?
Join The Figure User Group: teapawt.org/i/chat/group_join/g211…
teapawt.org/robixone/status/211360…
Figure's F.03 humanoid on a treadmill. CEO Brett Adcock says 9 straight days, 348 miles, 1.2M steps, to stress-test for failures. One fixed camera, no cuts; a cable runs from its waist to the floor. By @Figure_robot
What convinces you more: 9 days of walking or 1 day of chores?
37
Overlooked NEO post of the week: @RealAlphaApp's notes on 1X CEO Bernt Børnich's Relentless interview. 50,000 shipments targeted for 2027, quality gates, tendon drives and a 4-week CAD-to-walking loop.
Join The NEO Users Group: teapawt.org/i/chat/group_join/g200…
teapawt.org/RealAlphaApp/status/21…
In-house manufacturing lets 1X iterate in four weeks
Bernt Børnich on Relentless with Ti Morse. Bernt Børnich is the founder and CEO of 1X, a California-based humanoid robotics company. 1X developed NEO, its first home robot, and is preparing to manufacture and ship 50,000 units in 2027.
Børnich says 1X targets 50,000 robot shipments in 2027, but argues that quality gates, returns and field learning matter more than line capacity alone. He describes an in-house-designed robot and a four-week hardware iteration loop, with staged deployment and field service to incorporate customer feedback. On intelligence, he favors broad models trained across human, internet and robot data, arguing that deployment needs diverse experiences rather than repeated narrow tasks. He sees social behavior, warmth and playful exploration as part of a useful home robot, and says building a company requires enjoying continuous problem-solving.
Minute by minute
1. Quality gates shape the ramp (0:00)
Børnich says the Hayward factory is built for 10,000 units at full capacity, while 1X is building a San Carlos facility for 100,000 units per year. He says the goal is to ship 50,000 robots in 2027, but the company needs to ramp through quality gates because problems emerge at higher volumes and returned robots would undermine the target.
2. Humanoids face broader deployment (1:37)
Børnich compares a well-designed humanoid robot, at about 1,000 parts and hopefully under 70 pounds, with a car at about 50,000 parts and 4,000 pounds. He agrees that robots will encounter more unusual situations because they work in less constrained environments. He says 1X will manage expectations, support customers and use more structured settings for some deployments.
3. Home use makes safety central (3:44)
Børnich says he began testing a wheeled robot called Eve at home about three years ago, and that experience helped motivate the move to a bipedal robot for household tasks. He says progress is slow because new data, hardware changes and redeployment take time. He identifies safety as the key trust barrier and says 1X aims to prove safety more formally.
4. General models could cover tasks (5:43)
Børnich says expert operators can test tasks before the robot can automate them. Rather than automating each task separately, 1X aims to build a general model that can learn from human movement and internet video, though it may initially be less capable at specialized tasks. He describes a cycle in which deployed robots attempt tasks, successes and failures create data, and better models improve their performance.
5. In-house design speeds iteration (9:12)
Børnich says 1X's in-house motors and tendon drives make its system simpler, with fewer parts, tolerances and sensors, but the company has had to build its own supply chain. He says major hardware changes can go from CAD to a walking robot in about four weeks, after a decade of developing production processes. He also says early robots may be serviced or remanufactured in the field so fixes can reach existing units.
6. Simplicity defines machine beauty (13:19)
Børnich says he sees advanced engineering as art and finds beauty in simplicity rather than complexity. He cites reducing cables and connectors, and combining two motor-shaft components into one. His design principle is to identify the minimum complexity needed and justify any parts beyond it.
7. Small teams preserve system understanding (16:58)
Børnich says he would have kept the team smaller and leaner for longer, because more people require interfaces that can add complexity to the product. He values people who understand the machine's fundamentals across multiple domains. He says that deep understanding takes time to build, and that strong engineers often create tools to help them do their jobs.
8. Warmth depends on behavior too (20:13)
Børnich says 1X avoids intimidating, science-fiction styling and wants its technology to blend into society. He argues that world models could learn social behavior as well as physical tasks, such as holding a bag until someone takes it. He says robots may become personal companions, but that shaping their behavior will require balancing safety with personality and choosing data carefully.
9. Play could support exploration (25:46)
Børnich says he hopes robots become curious and playful, exploring things such as how sand moves. He thinks real-world experimentation can help a robot test its hypotheses, and says a developer platform could expose robots to more tasks and environments. He also suggests that rising expectations for household chores may leave robots with less idle time than expected.
10. Models need broad world data (28:51)
Børnich argues that models should learn how the world works rather than simply copy human behavior, then act according to the desired outcome. He compares this with driving. Broad data can teach a model about traffic and its consequences, rather than just cloning human drivers. He says broad models can be distilled for specific tasks, but narrow models cannot become general. He expects specialized models to remain better in 2026 and a shift toward more general intelligence in 2027.
11. Robot data needs diversity (34:59)
Børnich says 1X uses web data, simulation, sensor data, video, teleoperation and robots learning by doing. He argues that robotics is usually diversity-bound, not simply data-bound, and that deployments should cover varied environments and applications. 1X plans a mix of enterprise, home and platform deployments. Børnich says folding is now a comparatively simple AI task, despite being a difficult problem for classical robotics.
12. Founders must enjoy the process (41:23)
Børnich says disappointment comes every week and that solving one set of problems only reveals harder ones. He argues that founders need to enjoy the journey because there is no final point when everything is fixed. He says he prefers the highs and lows to the middle and did not start the company for the money.
The link in the reply plays each point from its second.
37
Overlooked NEO post of the week: @RealAlphaApp's notes on 1X CEO Bernt Børnich's Relentless interview. 50,000 shipments targeted for 2027, quality gates, tendon drives and a 4-week CAD-to-walking loop.
Join The NEO Users Group: teapawt.org/i/chat/group_join/g200…
teapawt.org/RealAlphaApp/status/21…
In-house manufacturing lets 1X iterate in four weeks
Bernt Børnich on Relentless with Ti Morse. Bernt Børnich is the founder and CEO of 1X, a California-based humanoid robotics company. 1X developed NEO, its first home robot, and is preparing to manufacture and ship 50,000 units in 2027.
Børnich says 1X targets 50,000 robot shipments in 2027, but argues that quality gates, returns and field learning matter more than line capacity alone. He describes an in-house-designed robot and a four-week hardware iteration loop, with staged deployment and field service to incorporate customer feedback. On intelligence, he favors broad models trained across human, internet and robot data, arguing that deployment needs diverse experiences rather than repeated narrow tasks. He sees social behavior, warmth and playful exploration as part of a useful home robot, and says building a company requires enjoying continuous problem-solving.
Minute by minute
1. Quality gates shape the ramp (0:00)
Børnich says the Hayward factory is built for 10,000 units at full capacity, while 1X is building a San Carlos facility for 100,000 units per year. He says the goal is to ship 50,000 robots in 2027, but the company needs to ramp through quality gates because problems emerge at higher volumes and returned robots would undermine the target.
2. Humanoids face broader deployment (1:37)
Børnich compares a well-designed humanoid robot, at about 1,000 parts and hopefully under 70 pounds, with a car at about 50,000 parts and 4,000 pounds. He agrees that robots will encounter more unusual situations because they work in less constrained environments. He says 1X will manage expectations, support customers and use more structured settings for some deployments.
3. Home use makes safety central (3:44)
Børnich says he began testing a wheeled robot called Eve at home about three years ago, and that experience helped motivate the move to a bipedal robot for household tasks. He says progress is slow because new data, hardware changes and redeployment take time. He identifies safety as the key trust barrier and says 1X aims to prove safety more formally.
4. General models could cover tasks (5:43)
Børnich says expert operators can test tasks before the robot can automate them. Rather than automating each task separately, 1X aims to build a general model that can learn from human movement and internet video, though it may initially be less capable at specialized tasks. He describes a cycle in which deployed robots attempt tasks, successes and failures create data, and better models improve their performance.
5. In-house design speeds iteration (9:12)
Børnich says 1X's in-house motors and tendon drives make its system simpler, with fewer parts, tolerances and sensors, but the company has had to build its own supply chain. He says major hardware changes can go from CAD to a walking robot in about four weeks, after a decade of developing production processes. He also says early robots may be serviced or remanufactured in the field so fixes can reach existing units.
6. Simplicity defines machine beauty (13:19)
Børnich says he sees advanced engineering as art and finds beauty in simplicity rather than complexity. He cites reducing cables and connectors, and combining two motor-shaft components into one. His design principle is to identify the minimum complexity needed and justify any parts beyond it.
7. Small teams preserve system understanding (16:58)
Børnich says he would have kept the team smaller and leaner for longer, because more people require interfaces that can add complexity to the product. He values people who understand the machine's fundamentals across multiple domains. He says that deep understanding takes time to build, and that strong engineers often create tools to help them do their jobs.
8. Warmth depends on behavior too (20:13)
Børnich says 1X avoids intimidating, science-fiction styling and wants its technology to blend into society. He argues that world models could learn social behavior as well as physical tasks, such as holding a bag until someone takes it. He says robots may become personal companions, but that shaping their behavior will require balancing safety with personality and choosing data carefully.
9. Play could support exploration (25:46)
Børnich says he hopes robots become curious and playful, exploring things such as how sand moves. He thinks real-world experimentation can help a robot test its hypotheses, and says a developer platform could expose robots to more tasks and environments. He also suggests that rising expectations for household chores may leave robots with less idle time than expected.
10. Models need broad world data (28:51)
Børnich argues that models should learn how the world works rather than simply copy human behavior, then act according to the desired outcome. He compares this with driving. Broad data can teach a model about traffic and its consequences, rather than just cloning human drivers. He says broad models can be distilled for specific tasks, but narrow models cannot become general. He expects specialized models to remain better in 2026 and a shift toward more general intelligence in 2027.
11. Robot data needs diversity (34:59)
Børnich says 1X uses web data, simulation, sensor data, video, teleoperation and robots learning by doing. He argues that robotics is usually diversity-bound, not simply data-bound, and that deployments should cover varied environments and applications. 1X plans a mix of enterprise, home and platform deployments. Børnich says folding is now a comparatively simple AI task, despite being a difficult problem for classical robotics.
12. Founders must enjoy the process (41:23)
Børnich says disappointment comes every week and that solving one set of problems only reveals harder ones. He argues that founders need to enjoy the journey because there is no final point when everything is fixed. He says he prefers the highs and lows to the middle and did not start the company for the money.
The link in the reply plays each point from its second.
37
Overlooked Optimus post of the week: @Kisalay_ reads Tesla's Q3 update closely. Fremont is in "Early Ramp" with the first production-quality bots, and Musk's timeline is measured. Worth a look.
Join The Optimus Users Group: teapawt.org/i/chat/group_join/g211…
teapawt.org/Kisalay_/status/211484…
Optimus is now coming off Tesla's high volume lines in Fremont.
Tesla's Q3 update, filed Oct 21, says production of the first production quality bots has begun there. Fremont is listed as Early Ramp. Texas is listed as Construction.
On the call, Musk placed the bulk of a 1 million bot per year line in Fremont.
His timeline was measured:
→ At least a year, maybe a little longer, before output hits 20,000 bots weekly
→ Roughly 10,000 things have to go right
→ Most issues should be solved around the end of next year, which is when outside deliveries come
The Verge noted that Tesla gave no unit count. Hands, safety, reliability, durability and the AI itself still need work.
All of this came in a quarter where Tesla profit fell 43 percent, per TechCrunch. The robot is being built while the core business is under pressure.
A production line is the point where a demo becomes a cost per unit.
A prototype impresses.
A line has to repeat.
2
62
If you're counting down to your own Optimus, The Optimus Users Group is for you. @California88658 @Teslaconomics @Maximilius31 @strengthPlan, we'd love your perspective in there: teapawt.org/i/chat/group_join/g211…
25












