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.

We pushed MolmoAct 2 into a completely out-of-distribution studio setting to test its robustness under extreme environmental changes. Despite only 10 minutes of task-specific fine-tuning data collected elsewhere, it adapted surprisingly well to the new environment.

Jiafei Duan

6,444 subscribers

11,462 views • 3 months ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

LLM post-training used to mean fine-tuning to a downstream task Robotics has been stuck in this setting, needing task-specific fine-tuning for best performance π07 changes this: It works out of the box & outperforms fine-tuned specialists Details:

LLM post-training used to mean fine-tuning to a downstream task Robotics has been stuck in this setting, needing task-specific fine-tuning for best performance π07 changes this: It works out of the box & outperforms fine-tuned specialists Details:

Chelsea Finn

68,106 views • 3 months ago

MolmoAct 2 builds on MolmoAct, our first Action Reasoning Model (ARM). Like its predecessor, MolmoAct 2 reasons about the world in 3D before taking actions. It now runs up to 37x faster & handles two-armed tasks out of the box without per-task fine-tuning.

MolmoAct 2 builds on MolmoAct, our first Action Reasoning Model (ARM). Like its predecessor, MolmoAct 2 reasons about the world in 3D before taking actions. It now runs up to 37x faster & handles two-armed tasks out of the box without per-task fine-tuning.

Ai2

25,563 views • 3 months ago

Sharing some exciting DYNA-1 result: zero-shot environment generalization We put DYNA-1 under test in a completely different environment from our training distribution – with an entirely different background (Dyna Robotics banner) and metal table. The table has a reflective and smooth surface, creating a wildly different visual appearance as well as interaction dynamics. The model is able to proceed as usual, adeptly folding and recovering from its own mistakes. By focusing on task mastery, we achieve robust generalization out of the box

Sharing some exciting DYNA-1 result: zero-shot environment generalization We put DYNA-1 under test in a completely different environment from our training distribution – with an entirely different background (Dyna Robotics banner) and metal table. The table has a reflective and smooth surface, creating a wildly different visual appearance as well as interaction dynamics. The model is able to proceed as usual, adeptly folding and recovering from its own mistakes. By focusing on task mastery, we achieve robust generalization out of the box

Jason Ma

49,008 views • 1 year ago

🚨 Scandalous 🚨 Nearly 90% of the plastic collected under the Deposit Return Scheme is shipped out of Ireland despite there being a company in Monaghan that’s able to recycle it. This is ludicrous for what’s supposed to be an environmental scheme. This must be addressed!

🚨 Scandalous 🚨 Nearly 90% of the plastic collected under the Deposit Return Scheme is shipped out of Ireland despite there being a company in Monaghan that’s able to recycle it. This is ludicrous for what’s supposed to be an environmental scheme. This must be addressed!

Matt Carthy TD

102,021 views • 1 year ago

New short course on Fine-tuning LLMs! Many developers are moving beyond only prompting, to also fine-tuning LLMs - that is, taking a pre-trained model and training it further on your own data, which can deliver superior results inexpensively. In this course, Sharon Zhou, CEO of Lamini (disclosure: I’m a minor shareholder) shows you how to recognize when fine-tuning can be help, and how to train an open-source LLM on your own data. I hope you enjoy the course!

New short course on Fine-tuning LLMs! Many developers are moving beyond only prompting, to also fine-tuning LLMs - that is, taking a pre-trained model and training it further on your own data, which can deliver superior results inexpensively. In this course, Sharon Zhou, CEO of Lamini (disclosure: I’m a minor shareholder) shows you how to recognize when fine-tuning can be help, and how to train an open-source LLM on your own data. I hope you enjoy the course!

Andrew Ng

502,854 views • 3 years ago

OpenAI engineer explained every form of fine-tuning in 1 hour 46 minutes - better than $3000 fine-tuning bootcamps. pick your task -> try prompting first -> drop to SFT for style -> run DPO for preference -> run RFT for verifiable rewards -> ship a model 10x cheaper than GPT-5. That loop is why teams that fine-tune beat teams that only prompt on cost and speed. Prompting + SFT + DPO + RFT + OpenAI fine-tuning API - that's the stack. Watch and save it, then pick the right fine-tune for your task this week.

OpenAI engineer explained every form of fine-tuning in 1 hour 46 minutes - better than $3000 fine-tuning bootcamps. pick your task -> try prompting first -> drop to SFT for style -> run DPO for preference -> run RFT for verifiable rewards -> ship a model 10x cheaper than GPT-5. That loop is why teams that fine-tune beat teams that only prompt on cost and speed. Prompting + SFT + DPO + RFT + OpenAI fine-tuning API - that's the stack. Watch and save it, then pick the right fine-tune for your task this week.

h100envy

40,941 views • 26 days ago

UPDATE on our S1 test anomaly! Our first stage is gone – but we have collected a significant amount of data and film footage. After reviewing it, we want to share some insights into the technical analysis, preliminary conclusions and our next steps ⬇️

UPDATE on our S1 test anomaly! Our first stage is gone – but we have collected a significant amount of data and film footage. After reviewing it, we want to share some insights into the technical analysis, preliminary conclusions and our next steps ⬇️

Rocket Factory Augsburg

167,193 views • 1 year ago

The end of 2025 marks the conclusion of an incredible two-year run for Automobilista 2: from January 2024 to December 2025, we delivered 43 new cars, 14 new classes, and 10 new tracks along with many updates to existing content and a wide-ranging catalogue of core gameplay and simulation improvements. These developments have brought in hundreds of thousands of new players to the base as we continue to expand AMS2 into new motorsport disciplines and eras. Our shared passion for motorsport and sim racing is what drives us and is what we celebrate with this trailer, as well as the relentless efforts of our team and the unwavering support of our players, which allow us to keep pushing forward and setting ever-higher goals with each new year. It is in that spirit that we head into 2026: committed not only to build upon the incredible foundation already established, but to elevate it into something greater than the sum of its parts. We wish all of you a Very Happy New Year!

The end of 2025 marks the conclusion of an incredible two-year run for Automobilista 2: from January 2024 to December 2025, we delivered 43 new cars, 14 new classes, and 10 new tracks along with many updates to existing content and a wide-ranging catalogue of core gameplay and simulation improvements. These developments have brought in hundreds of thousands of new players to the base as we continue to expand AMS2 into new motorsport disciplines and eras. Our shared passion for motorsport and sim racing is what drives us and is what we celebrate with this trailer, as well as the relentless efforts of our team and the unwavering support of our players, which allow us to keep pushing forward and setting ever-higher goals with each new year. It is in that spirit that we head into 2026: committed not only to build upon the incredible foundation already established, but to elevate it into something greater than the sum of its parts. We wish all of you a Very Happy New Year!

Reiza Studios

19,858 views • 7 months ago

Every Prufrock that leaves the factory has over 1,000 changes from the one before it. Before it ships, we put it through a full Porpoise Test (PT) in Bastrop — a short test tunnel (typically under 200 rings) that lets both the design team and the test team gather real data on the new designs and on the specific Prufrock serial number. To date, we’ve completed six of these tests and Prufrock 5 is currently 200 rings into PT7. Once complete, the tunnels are put to use: • PT4 is now a vacuum tunnel for Hyperloop testing • PT5 is a live test bench for liner trucks and new belt systems • PT6 is a pedestrian tunnel connecting SpaceX to Hyperloop Plaza With porpoise testing, our engineers get to get hands on and stress-test every one of their designs in real world environments without ever leaving Bastrop. We are currently hiring Electrical, Mechanical, Test, and Software engineers. Link to apply in the comments below 👇

Every Prufrock that leaves the factory has over 1,000 changes from the one before it. Before it ships, we put it through a full Porpoise Test (PT) in Bastrop — a short test tunnel (typically under 200 rings) that lets both the design team and the test team gather real data on the new designs and on the specific Prufrock serial number. To date, we’ve completed six of these tests and Prufrock 5 is currently 200 rings into PT7. Once complete, the tunnels are put to use: • PT4 is now a vacuum tunnel for Hyperloop testing • PT5 is a live test bench for liner trucks and new belt systems • PT6 is a pedestrian tunnel connecting SpaceX to Hyperloop Plaza With porpoise testing, our engineers get to get hands on and stress-test every one of their designs in real world environments without ever leaving Bastrop. We are currently hiring Electrical, Mechanical, Test, and Software engineers. Link to apply in the comments below 👇

Mike Baier

81,384 views • 5 days ago

I'm excited to share the new look of Agege Pen Cinema! Our team from the Ministry of the Environment and Water Resources has transformed this area into a model of environmental sustainability. I was delighted to be informed of a cool and calm atmosphere throughout the site - a direct result of our relentless efforts to maintain high standards of cleanliness and environmental management. A special shout-out to Comr. Olaide Oladeebo Okadigbo and his team, who have successfully zeroed-out the black spot along the railway track. Their dedication ensures we uphold a zero-tolerance policy for any practices that compromise our environment. Together, we’re making meaningful strides in enhancing public spaces and setting a benchmark for environmental excellence. I thank all our personnel for their hard work and commitment. Let’s continue working together to ensure that Lagos remains a beacon of cleanliness and sustainable development. #CleanerLagos #ZeroTolerance

I'm excited to share the new look of Agege Pen Cinema! Our team from the Ministry of the Environment and Water Resources has transformed this area into a model of environmental sustainability. I was delighted to be informed of a cool and calm atmosphere throughout the site - a direct result of our relentless efforts to maintain high standards of cleanliness and environmental management. A special shout-out to Comr. Olaide Oladeebo Okadigbo and his team, who have successfully zeroed-out the black spot along the railway track. Their dedication ensures we uphold a zero-tolerance policy for any practices that compromise our environment. Together, we’re making meaningful strides in enhancing public spaces and setting a benchmark for environmental excellence. I thank all our personnel for their hard work and commitment. Let’s continue working together to ensure that Lagos remains a beacon of cleanliness and sustainable development. #CleanerLagos #ZeroTolerance

Tokunbo Wahab

39,118 views • 1 year ago

Literally nobody in all of history has had a good time of it. Feminist are made because they couldn't vote until 10 minutes ago? Fine! Men couldn't vote until 20 minutes ago, and also we had to throw ourselves into the meat grinder. 🤨

Literally nobody in all of history has had a good time of it. Feminist are made because they couldn't vote until 10 minutes ago? Fine! Men couldn't vote until 20 minutes ago, and also we had to throw ourselves into the meat grinder. 🤨

Fight With Memes

185,640 views • 8 months ago

Training humanoid robots on teleoperation alone won’t scale, says Rhoda AI CEO Jagdeep Singh. Unlike self-driving cars - which basically have four actuators (left, right, speed up, and slow down) and operate in a single environment (the road) - humanoid robots are completely different: "You're dealing with the full dexterity of a human hand - 20 degrees of freedom per hand. Every object is different. Every type of task is different." The problem isn't just the quantity of data - the bigger issue is diversity in the data, and it's why many humanoid robot demos will struggle to adapt to the real world: "If all the data you have is data that you've intentionally collected, then you almost by definition haven't seen the corner cases. You haven't seen all those edge scenarios that cause failure." Jagdeep says teleoperation is useful for fine-tuning robot behavior, but for pretraining, it's completely inadequate.

Training humanoid robots on teleoperation alone won’t scale, says Rhoda AI CEO Jagdeep Singh. Unlike self-driving cars - which basically have four actuators (left, right, speed up, and slow down) and operate in a single environment (the road) - humanoid robots are completely different: "You're dealing with the full dexterity of a human hand - 20 degrees of freedom per hand. Every object is different. Every type of task is different." The problem isn't just the quantity of data - the bigger issue is diversity in the data, and it's why many humanoid robot demos will struggle to adapt to the real world: "If all the data you have is data that you've intentionally collected, then you almost by definition haven't seen the corner cases. You haven't seen all those edge scenarios that cause failure." Jagdeep says teleoperation is useful for fine-tuning robot behavior, but for pretraining, it's completely inadequate.

TBPN

19,218 views • 5 months 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

11,514 views • 1 month ago

been building an agent-native, node-based, real-time environment. as a test, I gave gpt-5.6 sol a prompt to plan out its own setup and go into a loop of creative exploration until it felt that it was done. it invented every node, every data type, every connection and all of it comes out modular and human-readable. i could open any piece and steer while it built

been building an agent-native, node-based, real-time environment. as a test, I gave gpt-5.6 sol a prompt to plan out its own setup and go into a loop of creative exploration until it felt that it was done. it invented every node, every data type, every connection and all of it comes out modular and human-readable. i could open any piece and steer while it built

Spencer Sterling

22,217 views • 19 days ago

In my experience, robot 'generalists' are often jacks of all trades but masters of none. In training across multiple tasks and environments, robot policies fail to generalize robustly and effectively to each particular test setting. What if at test time, we non-parametrically *retrieved* “relevant” data from the training set and used it to significantly improve the performance of few-shot imitation learning to be robust to various test time scenes. Notably, we are *not* collecting lots of new data, just training more on sub-components of the same training data! Now, we’re certainly not the first to suggest retrieval, but in our new work - STRAP, we show how retrieving relevant *sub-trajectories* from offline datasets can significantly increase data reuse across tasks, when paired with an appropriate metric space. A 🧵 (1/7)

In my experience, robot 'generalists' are often jacks of all trades but masters of none. In training across multiple tasks and environments, robot policies fail to generalize robustly and effectively to each particular test setting. What if at test time, we non-parametrically retrieved “relevant” data from the training set and used it to significantly improve the performance of few-shot imitation learning to be robust to various test time scenes. Notably, we are not collecting lots of new data, just training more on sub-components of the same training data! Now, we’re certainly not the first to suggest retrieval, but in our new work - STRAP, we show how retrieving relevant sub-trajectories from offline datasets can significantly increase data reuse across tasks, when paired with an appropriate metric space. A 🧵 (1/7)

Abhishek Gupta

12,045 views • 1 year ago

Pulp Fiction told its story out of order in 1994 and trusted a Friday night crowd to keep the pieces straight. It worked so well that scrambling the timeline stopped being an art house move and turned into a default setting.

Pulp Fiction told its story out of order in 1994 and trusted a Friday night crowd to keep the pieces straight. It worked so well that scrambling the timeline stopped being an art house move and turned into a default setting.

Quentin Tarantino Universe

175,140 views • 12 days ago

Learned visuomotor policies are notoriously fragile, they break with changes in conditions like lighting, clutter, or object variations amongst other things. In Yunchu @ CoRL2025's latest work, we asked whether we could get these policies to be robust and generalizable with a clever choice of visual representation! The argument we made was - we want a choice of visual representation that specifically adapts to be sufficient, yet minimal for the task at hand. We thought about it from the perspective of flexible, key-point based representations. The key question becomes - how do we choose a sufficient, task-specific, yet minimal set of keypoints as a representation for policy learning. Yunchu proposes a neat way of automatically selecting task-relevant keypoints using a standard supervised learning objective, and using this for robust policy learning. This is largely under the same assumptions as behavior cloning, but with huge gains on robustness. Let’s understand how, 🧵 (1/8)

Learned visuomotor policies are notoriously fragile, they break with changes in conditions like lighting, clutter, or object variations amongst other things. In Yunchu @ CoRL2025's latest work, we asked whether we could get these policies to be robust and generalizable with a clever choice of visual representation! The argument we made was - we want a choice of visual representation that specifically adapts to be sufficient, yet minimal for the task at hand. We thought about it from the perspective of flexible, key-point based representations. The key question becomes - how do we choose a sufficient, task-specific, yet minimal set of keypoints as a representation for policy learning. Yunchu proposes a neat way of automatically selecting task-relevant keypoints using a standard supervised learning objective, and using this for robust policy learning. This is largely under the same assumptions as behavior cloning, but with huge gains on robustness. Let’s understand how, 🧵 (1/8)

Abhishek Gupta

11,355 views • 1 year ago

In our situation at the U.S. Agency for Global Media, we're trying to streamline, modernize, & return it to its actual mission. But we have a federal judge intervening to dictate personnel & how we operate. It's completely out of control. These judges are not only usurping the power of the President but subverting the will of the people.

In our situation at the U.S. Agency for Global Media, we're trying to streamline, modernize, & return it to its actual mission. But we have a federal judge intervening to dictate personnel & how we operate. It's completely out of control. These judges are not only usurping the power of the President but subverting the will of the people.

Kari Lake

110,876 views • 1 year ago

another 100% AI podcast > one prompt > 2 minutes from zero to final video > no time limit - make it as long as you need > can be adapted to any script, character, and environment you need a lot of you have been asking how we automated all these videos, we’re launching the beta very soon the real UGC industry isn't dying, it already died

another 100% AI podcast > one prompt > 2 minutes from zero to final video > no time limit - make it as long as you need > can be adapted to any script, character, and environment you need a lot of you have been asking how we automated all these videos, we’re launching the beta very soon the real UGC industry isn't dying, it already died

MAX

30,603 views • 4 months ago