Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

🤖What if a robot could perform a new task just from a natural language command, with zero demonstrations? Our new work, NovaFlow, makes it possible! We use pre-trained video generative model to create a video of the task, then translate it into a plan for real-world robot execution. 1/6... show more

Hongyu Li

1,948 subscribers

106,422 views • 1 year ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

34 Comments

Hongyu Li1 year ago

We start by giving a video model a text prompt like "hang the mug." It generates a short clip of a human performing the task, acting as a visual instruction manual. This video provides the commonsense physical knowledge needed for the robot to understand the motion. 🎥 2/6

Hongyu Li1 year ago

The generated video is 2D, so how does the robot understand 3D space? We lift it! 📏 A monocular video depth estimation model predicts depth for every frame. Then, we calibrate it with the robot's real-world depth camera for precise spatial awareness. 3/6

Hongyu Li1 year ago

🤔 But how do you bridge the gap between that video and physical control? We introduce "actionable 3D object flow," a novel representation distilled from the video that guides the robot’s motion for complex rigid, articulated, and even deformable objects. 🌟 4/6

Hongyu Li1 year ago

NovaFlow shows incredible generalization. From rigid blocks to deformable ropes, on both a tabletop arm and a mobile Spot! It adapts to new objects and embodiments without any retraining. It achieves SOTA zero-shot results, even outperforming policies trained on 10-30 demos. 🏆 5/6

Hongyu Li1 year ago

By leveraging the commonsense knowledge in video models, NovaFlow helps solve the robot data bottleneck. This is a leap towards more general-purpose robots. 🚀 Dive deeper into our work: 📜 Paper: 🌐 Project Website: 👀 P.S. You can check out the 3D generated videos and planned robot trajectories in a live viewer on our website! 6/6

Hongyu Li1 year ago

This is a collaboration with @ling_feng_sun, @YafeiHuCMU, Duy Ta, Jenny Barry, and @jiahuifu_carol.

Nils1 year ago

Beautiful.

Hongyu Li1 year ago

Thanks, Nils!

Xiatao Sun1 year ago

Using video generation models to create task plans from language is clever! The actionable 3D object flow representation seems like the key to making this work across embodiments.

Hongyu Li1 year ago

Thanks, Xiatao. I believe video models capture the most general form of dynamics, although they can’t be directly understood by robots. Actionable flow enables robots to comprehend the generated videos.

Zhiyang (Frank) Dou1 year ago

Nice work!

Hongyu Li1 year ago

Thanks, Zhiyang!

Bozheng Li1 year ago

great work!

Hongyu Li1 year ago

Thanks, Bozheng!

Balveer Singh1 year ago

Wow, this is such a cool approach.

Hongyu Li1 year ago

Thanks🙏

Ted Xiao1 year ago

Nice work!

Hongyu Li1 year ago

Thanks, Ted! 🙏

Sam Green1 year ago

Laying the groundwork for Soma

Atharva Sawant1 year ago

So cool

Hongyu Li1 year ago

Thanks

Cryptlesh1 year ago

Wow. It’s like imagining whole take on head and then implementing it like human break does. Kind of.

Hongyu Li1 year ago

Exactly, a video model serves as the imagination capabilities of human beings.

Bercan1 year ago

Great work. Can you try on these examples in mcap.

Hongyu Li1 year ago

Thanks for the great suggestions! Egocentric data is more challenging, and we’re still working on it!

Bercan1 year ago

Happy to work together we are also working on real2sim

Stanley Wei1 year ago

That will save a lot of time. Hope you post some cool demos.

BlockChan11 months ago

holy shit i personaly owned name for a while and have no idea what to build

towerofshadow1 year ago

is it more effective to train robots using human actions that are tracked using actionable flow, or to teleoperate robots into direct movement data?

Hongyu Li1 year ago

NovaFlow is object-centric and embodiment-agnostic. So it should work the same for both human and robot data. In practice, we find video generative models work better with generating human data.

Danny Ng1 year ago

I wonder how long it takes to perform one task?

Hongyu Li1 year ago

It depends on the choice of video model. For Veo, it takes around 2-3min for the entire pipeline.

Urs Gehrig1 year ago

Zero shot Robotic AI manipulation.

Mariano Phielipp8 months ago

Interesting work. How you see the work of Florence, Manuelli, Redrake in the context of this work Dense Object Nets: Learning Dense Visual Object Descriptors By and For Robotic Manipulation Florence,Manuelli, Russ Tedrake?

Related Videos

We taught a robot to play! Introducing Dexterity from Supervised Play (DSP): turning minutes of human play into a world model and dexterous manipulation policy. No task demonstrations or sim2real!

We taught a robot to play! Introducing Dexterity from Supervised Play (DSP): turning minutes of human play into a world model and dexterous manipulation policy. No task demonstrations or sim2real!

Ruoshi Liu

26,153 views • 12 days ago

S1 learns new tasks like a language model. You prompt it with a video demonstration, and it outputs robot actions to complete the task in any environment and in any embodiment.

S1 learns new tasks like a language model. You prompt it with a video demonstration, and it outputs robot actions to complete the task in any environment and in any embodiment.

Skild AI

308,754 views • 2 months ago

World Labs co-founder Justin Johnson predicts a real-to-sim-to-real future where Atlas could train a robot for any new environment in minutes using just 5 photos from your phone: "One is this notion of fully in-context learning. Maybe I've got a robotics foundation model, then I can demonstrate the robot once how a task should be performed, and that's enough for the robot to figure out." "Another version is real to sim to real. Maybe I've got my pre-trained robotics foundation model, but I want to adapt it to this particular environment, like this studio, this table, moving this microphone around." "There's a version where you could come in here and take just a couple casual videos with your phone or a couple images with your phone, use Atlas to reconstruct the space, and now stage all kinds of robotics interactions in this studio space." "If you could lower the barrier to doing that, you could do it in five minutes to take five photos, upload it to Atlas, have it generate a simulation, maybe describe in natural language what kind of task you want the robot to do, have an agent go build that simulation for you, RL fine-tune your general purpose robotics foundation model, and now you've got a robot maybe in the span of a couple minutes that could come in and is perfectly adapted to this space." World Labs

World Labs co-founder Justin Johnson predicts a real-to-sim-to-real future where Atlas could train a robot for any new environment in minutes using just 5 photos from your phone: "One is this notion of fully in-context learning. Maybe I've got a robotics foundation model, then I can demonstrate the robot once how a task should be performed, and that's enough for the robot to figure out." "Another version is real to sim to real. Maybe I've got my pre-trained robotics foundation model, but I want to adapt it to this particular environment, like this studio, this table, moving this microphone around." "There's a version where you could come in here and take just a couple casual videos with your phone or a couple images with your phone, use Atlas to reconstruct the space, and now stage all kinds of robotics interactions in this studio space." "If you could lower the barrier to doing that, you could do it in five minutes to take five photos, upload it to Atlas, have it generate a simulation, maybe describe in natural language what kind of task you want the robot to do, have an agent go build that simulation for you, RL fine-tune your general purpose robotics foundation model, and now you've got a robot maybe in the span of a couple minutes that could come in and is perfectly adapted to this space." World Labs

MTS

25,204 views • 1 month ago

Xiaomi Robotics-1 🤖a new robot foundation model from Xiaomi Tech - 100K+ hours real world robot data - VLA architecture - Cross-embodiment generalization - Fast adaptation to new tasks

Xiaomi Robotics-1 🤖a new robot foundation model from Xiaomi Tech - 100K+ hours real world robot data - VLA architecture - Cross-embodiment generalization - Fast adaptation to new tasks

Adina Yakup

17,380 views • 2 months ago

World Labs co-founder Justin Johnson on how five photos from your phone could teach a robot to work in a room it has never seen: "I think there's a couple different tech trees that people are working on. One is this notion of fully in-context learning. Maybe I've got a robotics foundation model, then I can demonstrate to the robot once how a task should be performed, and that's enough for the robot to figure out that task." "Another version is what we're calling real-to-sim-to-real. Maybe I've got my pre-trained robotics foundation model, but I want to adapt it to this particular environment." "Like this studio... You could come in here and take just a couple casual videos with your phone or a couple images with your phone, use Atlas to reconstruct the space, and now stage all kinds of robotics interactions in this studio space in particular." "Literally you could do it in five minutes. Take five photos, upload it to Atlas, have it generate a simulation, describe in natural language what kind of task you want the robot to do, have an agent go build that simulation for you, RL fine-tune your general purpose robotics foundation model, and now you've got a robot... that's perfectly adapted to this space." Justin Johnson on MTS

World Labs co-founder Justin Johnson on how five photos from your phone could teach a robot to work in a room it has never seen: "I think there's a couple different tech trees that people are working on. One is this notion of fully in-context learning. Maybe I've got a robotics foundation model, then I can demonstrate to the robot once how a task should be performed, and that's enough for the robot to figure out that task." "Another version is what we're calling real-to-sim-to-real. Maybe I've got my pre-trained robotics foundation model, but I want to adapt it to this particular environment." "Like this studio... You could come in here and take just a couple casual videos with your phone or a couple images with your phone, use Atlas to reconstruct the space, and now stage all kinds of robotics interactions in this studio space in particular." "Literally you could do it in five minutes. Take five photos, upload it to Atlas, have it generate a simulation, describe in natural language what kind of task you want the robot to do, have an agent go build that simulation for you, RL fine-tune your general purpose robotics foundation model, and now you've got a robot... that's perfectly adapted to this space." Justin Johnson on MTS

a16z

35,947 views • 1 month ago

Robotics is moving beyond teaching robots how to perform a fixed set of actions. nature will’s new foundation model is built around a more general problem: can a robot understand an instruction, adapt to the situation in front of it, and physically execute the task? What makes this interesting is the combination of dexterous control, zero-shot generalization and long-horizon execution in one model.

Robotics is moving beyond teaching robots how to perform a fixed set of actions. nature will’s new foundation model is built around a more general problem: can a robot understand an instruction, adapt to the situation in front of it, and physically execute the task? What makes this interesting is the combination of dexterous control, zero-shot generalization and long-horizon execution in one model.

Fathela Esq

85,798 views • 15 days ago

Most robot demos are scripted. Generalist's GEN-1 is not. > GEN-1 was doing a task > Mid-task, an extra object was thrown into the bin > GEN-1 quickly adapted to the new environment > Still completed the task smoothly This is the difference between a scripted policy (seen in viral launch videos) and a real foundation model. GEN-1 was trained on 500,000 hours of proprietary, real-world dexterous manipulation data. It has NEVER seen that exact scenario before. This was the coolest demo at Automate Show Generalist Flexiv Robotics

Most robot demos are scripted. Generalist's GEN-1 is not. > GEN-1 was doing a task > Mid-task, an extra object was thrown into the bin > GEN-1 quickly adapted to the new environment > Still completed the task smoothly This is the difference between a scripted policy (seen in viral launch videos) and a real foundation model. GEN-1 was trained on 500,000 hours of proprietary, real-world dexterous manipulation data. It has NEVER seen that exact scenario before. This was the coolest demo at Automate Show Generalist Flexiv Robotics

mete

11,651 views • 4 months ago

Xiaomi-Robotics-1 just dropped on Hugging Face 🔥 A robot foundation model trained on 100,000 hours of real-world manipulation. They turned it loose in a real apartment: folding laundry, loading the washer, doing the dishes, packing a suitcase. Fully autonomous.

Xiaomi-Robotics-1 just dropped on Hugging Face 🔥 A robot foundation model trained on 100,000 hours of real-world manipulation. They turned it loose in a real apartment: folding laundry, loading the washer, doing the dishes, packing a suitcase. Fully autonomous.

Victor M

131,376 views • 3 months ago

How well do egocentric human videos scale? Apparently, really well. Dyna Robotics has introduced Dyna-2, a world-action model (WAM) pre-trained on more than 1 million hours of real egocentric human video. It learns by jointly predicting the next video frames and the actions, and that video prediction (world modeling) is what carries learning from humans to robots. • More human video predictably improves action prediction. • Task accuracy improves across embodiments (bimanual arms and a humanoid) with zero robot data in pre-training. • Post-trained on just a few hours of robot data, mean task performance rose 20% to 53% as human pre-training scaled 1k to 1M hours. WAM vs VLA, matched apple-to-apple: the WAM hit 1.55x the VLA's success rate and won 65% of head-to-heads. Deployed zero-shot at unseen customer sites, it passed 87% vs the VLA's 46%.

How well do egocentric human videos scale? Apparently, really well. Dyna Robotics has introduced Dyna-2, a world-action model (WAM) pre-trained on more than 1 million hours of real egocentric human video. It learns by jointly predicting the next video frames and the actions, and that video prediction (world modeling) is what carries learning from humans to robots. • More human video predictably improves action prediction. • Task accuracy improves across embodiments (bimanual arms and a humanoid) with zero robot data in pre-training. • Post-trained on just a few hours of robot data, mean task performance rose 20% to 53% as human pre-training scaled 1k to 1M hours. WAM vs VLA, matched apple-to-apple: the WAM hit 1.55x the VLA's success rate and won 65% of head-to-heads. Deployed zero-shot at unseen customer sites, it passed 87% vs the VLA's 46%.

The Humanoid Hub

15,988 views • 2 months ago

The Real-to-Sim part of R2S2R transforms physical robots, sensors, environments, objects, and interactions into simulations that preserve task-relevant observations and dynamics - not only how the world looks, but how it acts when the robot interacts with it. These results set a new bar for sim-real alignment in robot manipulation.

The Real-to-Sim part of R2S2R transforms physical robots, sensors, environments, objects, and interactions into simulations that preserve task-relevant observations and dynamics - not only how the world looks, but how it acts when the robot interacts with it. These results set a new bar for sim-real alignment in robot manipulation.

Fei-Fei Li

26,398 views • 3 months ago

Qwen-VLA feels like one of the first real robotics foundation models. A single system trained across robot manipulation, navigation, egocentric human video, simulation, and vision-language reasoning instead of isolated robot policies.

Qwen-VLA feels like one of the first real robotics foundation models. A single system trained across robot manipulation, navigation, egocentric human video, simulation, and vision-language reasoning instead of isolated robot policies.

Robots Digest 🤖

14,722 views • 5 months ago

Announcing RoboFuME🤖💨, a system for autonomous & efficient real-world robot learning that 1. pre-trains a VLM reward model and a multi-task RL policy from diverse off-the-shelf demo data; 2. runs RL fine-tuning online with the VLM reward model. 🔗 🧵↓

Announcing RoboFuME🤖💨, a system for autonomous & efficient real-world robot learning that 1. pre-trains a VLM reward model and a multi-task RL policy from diverse off-the-shelf demo data; 2. runs RL fine-tuning online with the VLM reward model. 🔗 🧵↓

Jingyun Yang

88,312 views • 3 years ago

The robot flipped a pancake nobody taught it! 🥞 Skild AI team assumed pancake flipping had to be somewhere in the training data. So they searched. Millions of hours of pre-training data. Nothing. S1 inferred the whole task from a single human demonstration. That's their new general robot model, built as an in-context learner from the ground up. Every new robot task today starts with days of teleoperation and a fine-tuning run on a specialist policy. S1 skips all of it. Much like a language model, it never updates its weights to learn a new task. The demonstration enters the context window, and the policy uses it to decide what to do next. → Ten-minute tasks it was never trained on, composed from primitives learned in pre-training: a new style of coffee, potting a plant, frying pancakes. → Soil and pots arrived at their office at 8:54 PM. The robot was running the task autonomously by 9:27 PM. → Slide objects away mid-reach, swap them, change the lighting, it still finishes. → The prompt waters a plant with a watering can, but only a cup is available. It uses the cup. It doesn't rigidly replay what it saw, but it recovers from its own errors, and sometimes executes with more precision than the demonstrator, when the human fumbles an egg and makes a mess, S1 performs the same step cleanly. The demonstration is a specification of the goal, and not a trajectory to copy. On unseen tasks after 100K hours of pre-training: language-prompted VLAs reach 9%. Their new model reaches 66%. It's already deploying with industrial partners, with a wider rollout over the coming months. Congrats Deepak Pathak and team behind this! 😮‍💨 🔗 Link to their latest blog: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

The robot flipped a pancake nobody taught it! 🥞 Skild AI team assumed pancake flipping had to be somewhere in the training data. So they searched. Millions of hours of pre-training data. Nothing. S1 inferred the whole task from a single human demonstration. That's their new general robot model, built as an in-context learner from the ground up. Every new robot task today starts with days of teleoperation and a fine-tuning run on a specialist policy. S1 skips all of it. Much like a language model, it never updates its weights to learn a new task. The demonstration enters the context window, and the policy uses it to decide what to do next. → Ten-minute tasks it was never trained on, composed from primitives learned in pre-training: a new style of coffee, potting a plant, frying pancakes. → Soil and pots arrived at their office at 8:54 PM. The robot was running the task autonomously by 9:27 PM. → Slide objects away mid-reach, swap them, change the lighting, it still finishes. → The prompt waters a plant with a watering can, but only a cup is available. It uses the cup. It doesn't rigidly replay what it saw, but it recovers from its own errors, and sometimes executes with more precision than the demonstrator, when the human fumbles an egg and makes a mess, S1 performs the same step cleanly. The demonstration is a specification of the goal, and not a trajectory to copy. On unseen tasks after 100K hours of pre-training: language-prompted VLAs reach 9%. Their new model reaches 66%. It's already deploying with industrial partners, with a wider rollout over the coming months. Congrats Deepak Pathak and team behind this! 😮‍💨 🔗 Link to their latest blog: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

Lukas Ziegler

11,406 views • 2 months ago

The robot is learning several novel tasks instantly, after just ONE demonstration each... Instant Policy makes it possible: no extra training, no weight updates, just pure in-context learning. It just got accepted at ICLR 2025, and it’s changing how robots learn. With just a single demo, a robot can pick up a new task and start performing it right away. Why this is a big deal: ✅ Learns tasks instantly with just one or a few demonstrations ✅ Improves over time as more demonstrations are given ✅ Uses simulation-based training with “pseudo-demonstrations” for scalability ✅ Can transfer skills across different robots and even follow language-defined tasks It brings in-context learning to robotics, opening up new possibilities for flexible, real-world automation. You can try it yourself: code and weights are available at • • • • Thank you to Edward Johns, Director of the Robot Learning Lab at Imperial College for sharing their work! 🙏

The robot is learning several novel tasks instantly, after just ONE demonstration each... Instant Policy makes it possible: no extra training, no weight updates, just pure in-context learning. It just got accepted at ICLR 2025, and it’s changing how robots learn. With just a single demo, a robot can pick up a new task and start performing it right away. Why this is a big deal: ✅ Learns tasks instantly with just one or a few demonstrations ✅ Improves over time as more demonstrations are given ✅ Uses simulation-based training with “pseudo-demonstrations” for scalability ✅ Can transfer skills across different robots and even follow language-defined tasks It brings in-context learning to robotics, opening up new possibilities for flexible, real-world automation. You can try it yourself: code and weights are available at • • • • Thank you to Edward Johns, Director of the Robot Learning Lab at Imperial College for sharing their work! 🙏

Ilir Aliu

46,285 views • 1 year ago

Humanoid robotics is entering a new phase： Xiaomi Robotics 0 (XR-0) explores learning manipulation directly from large-scale human videos — aligning video understanding with robot embodiment to generate executable actions. Watch humans->learn representations ->map to robot control. Robot data is scarce. Internet video is not. If video pretraining transfers reliably into embodied control, the scaling law of robotics changes. Deployment speed becomes a function of model alignment, not just hardware iteration. We’re moving from engineered skills to learned behavior. That’s a structural shift. project:

Humanoid robotics is entering a new phase： Xiaomi Robotics 0 (XR-0) explores learning manipulation directly from large-scale human videos — aligning video understanding with robot embodiment to generate executable actions. Watch humans->learn representations ->map to robot control. Robot data is scarce. Internet video is not. If video pretraining transfers reliably into embodied control, the scaling law of robotics changes. Deployment speed becomes a function of model alignment, not just hardware iteration. We’re moving from engineered skills to learned behavior. That’s a structural shift. project:

CyberRobo

18,542 views • 8 months ago

20+ household tasks. One foundation model. All on a real robot. Astribot just dropped one of the broadest real-robot household demos I’ve seen. Lumo-2 spans dexterous manipulation, physical understanding, temporal reasoning, human–robot collaboration, and long-horizon execution — all validated on a real robot across 20+ complex household tasks. That matters because Astribot has taken a full-stack approach from day one: AI, OS, and robot hardware developed together, with each demo showing how model capabilities translate into physical performance. Not a one-skill demo. One model pushed across the messy reality of home robotics. 🧵

20+ household tasks. One foundation model. All on a real robot. Astribot just dropped one of the broadest real-robot household demos I’ve seen. Lumo-2 spans dexterous manipulation, physical understanding, temporal reasoning, human–robot collaboration, and long-horizon execution — all validated on a real robot across 20+ complex household tasks. That matters because Astribot has taken a full-stack approach from day one: AI, OS, and robot hardware developed together, with each demo showing how model capabilities translate into physical performance. Not a one-skill demo. One model pushed across the messy reality of home robotics. 🧵

RoboHub🤖

34,367 views • 3 months ago

Can robot foundation models scale? Xiaomi-Robotics-1 explores this question with over 100,000 hours of real-world manipulation data. Pre-trained on large-scale real-world trajectories and post-trained with cross-embodiment robot data, Xiaomi-Robotics-1 demonstrates consistent scaling across data and model size, strong generalization in unseen environments, and efficient adaptation to new tasks. 🔗 #Robotics #EmbodiedAI #FoundationModels #RobotLearning #XiaomiAI #XiaomiRobotics

Can robot foundation models scale? Xiaomi-Robotics-1 explores this question with over 100,000 hours of real-world manipulation data. Pre-trained on large-scale real-world trajectories and post-trained with cross-embodiment robot data, Xiaomi-Robotics-1 demonstrates consistent scaling across data and model size, strong generalization in unseen environments, and efficient adaptation to new tasks. 🔗 #Robotics #EmbodiedAI #FoundationModels #RobotLearning #XiaomiAI #XiaomiRobotics

Xiaomi Tech

23,920 views • 3 months ago

How do you know your robot actually succeeded? In v0.6.0, you ask a VLM 🤖 LeRobot ships a new unified reward models API with two zero-shot additions. Robometer is a general-purpose reward model trained on 1M+ trajectories across 21 robot embodiments: point it at any LeRobot dataset and it scores task progress and success straight from video and a language instruction, no task-specific training needed. TOPReward goes even further: no reward weights at all. It wraps an off-the-shelf VLM and reads the log-probability of the token "True" given the video and the instruction. Both come with labeling scripts that write per-frame progress curves into your dataset, ready for reward-aware behavior cloning. More on how they work:

How do you know your robot actually succeeded? In v0.6.0, you ask a VLM 🤖 LeRobot ships a new unified reward models API with two zero-shot additions. Robometer is a general-purpose reward model trained on 1M+ trajectories across 21 robot embodiments: point it at any LeRobot dataset and it scores task progress and success straight from video and a language instruction, no task-specific training needed. TOPReward goes even further: no reward weights at all. It wraps an off-the-shelf VLM and reads the log-probability of the token "True" given the video and the instruction. Both come with labeling scripts that write per-frame progress curves into your dataset, ready for reward-aware behavior cloning. More on how they work:

LeRobot

22,592 views • 3 months ago

Show it a single human demonstration of a long-horizon task and the robot performs it right away. No fine-tuning, no retraining. Skild has released S1, a robot foundation model where you specify the task with a video demonstration instead of a language instruction. Why video, not language Language works for "hand me the mug." It falls apart for anything delicate or long-horizon. You don't learn to fold a fitted sheet from a sentence. So S1 takes a video demo into its context window and translates it to its own body and scene. What it does - Tasks up to 10 minutes, dozens of sequential steps - Unseen tasks: potting plants, pour-over coffee, kit assembly, pancakes. Skild says even the underlying primitives, like flipping a pancake, weren't in pre-training - Recovers from its own mistakes without being trained to - Improves on flawed demos. Demonstrator drops an egg, S1 doesn't make that mistake. It reads the demo as a goal, not a trajectory to copy.

Show it a single human demonstration of a long-horizon task and the robot performs it right away. No fine-tuning, no retraining. Skild has released S1, a robot foundation model where you specify the task with a video demonstration instead of a language instruction. Why video, not language Language works for "hand me the mug." It falls apart for anything delicate or long-horizon. You don't learn to fold a fitted sheet from a sentence. So S1 takes a video demo into its context window and translates it to its own body and scene. What it does - Tasks up to 10 minutes, dozens of sequential steps - Unseen tasks: potting plants, pour-over coffee, kit assembly, pancakes. Skild says even the underlying primitives, like flipping a pancake, weren't in pre-training - Recovers from its own mistakes without being trained to - Improves on flawed demos. Demonstrator drops an egg, S1 doesn't make that mistake. It reads the demo as a goal, not a trajectory to copy.

The Humanoid Hub

19,063 views • 2 months ago