What's different between these two BC policies? It's the... same architecture, training budget, and data collection setup — the only difference is the controller gains! Controller gains are an understudied design parameter in robot learning. In our new work (w/ antonia bronars*, Pulkit Agrawal), we show how they act as an inductive bias across BC, RL, and Sim2Real transfer, with real consequences on performance. Here's what we found 🧵 * Equal Contribution 📄arxiv: 🔗website:show more

Here’s a pretty weird and surprising result - retrieval-augmented... generation works unreasonably well for robot learning – but only when parameterized using difference vectors! We introduce Difference-Aware Retrieval Policies for Imitation Learning (DARP), a simple, semi-parametric RAG architecture for imitation learning that achieves gains of up to 200% over standard behavior cloning. No additional assumptions beyond BC, just a little architecture switch! The theory backing it up is pretty cool too and it works on real robots! :) Play with our website to understand better: 🧵(1/7)show more

How can we more effectively leverage robot data from... different embodiments for skill transfer? Excited to share that our new work, RoVi-Aug, has been accepted to Conference on Robot Learning as an oral paper！ WIth RoVi-Aug, you can augment an existing robot dataset into a different robot and different viewpoints. A policy trained on the augmented dataset can zero-shot deploy on the unseen target robot with significantly different camera angles! 🧵👇 🔗 Check out our paper:show more

Inkling-small is out today! With SGLang, you can get... 648 tok/s decode with DSpark (simulated acc len=4) and 288 tok/s w/o DSpark, under the same setup (8x NVIDIA AI B200, TP 8, NVFP4, bs=1). What makes this model different is the size. 276B total with 12B active is a sweet spot for RL, and both LoRA and full-parameter training become well within reach. Miles is ready and verified for multimodal RL on Inkling-small, so you can turn your multimodal data into real capability gains. At ~1/4 the size, Inkling-small matches the bigger version in capability and even wins on some benchmarks. Run Inkling-small with SGLang, and customize it with Miles.show more

Robora Sim: A PyBullet-Powered Environment for Learning Robotic Physical... Intelligence We are currently building our Robora simulation environment setup for our sim based learning, leveraging PyBullet, an industry-standard physics engine widely used in AI-driven robotics research and development. The environment is optimized with GPU-accelerated learning algorithms, enabling high-speed imitation learning and reinforcement learning within a safe and controlled virtual setup before shipping out to real world. This simulation platform allows our models to learn, adapt, and generalize across different robot morphologies, terrain types and task objectives - all before deployment to the real world. At it's core, the system combines a VLA-powered high-level planner with low-level motion control algorithms, working cohesively to produce emergent, physically intelligent behaviors. This synergy between simulation, learning, and real-world transfer marks a major step forward in our pursuit of adaptive and intelligent robotic systems. Through advanced domain randomization and synthetic data generation, the Robora Simulation Environment ensures that policies trained in simulation transfer effectively to real-world robots, minimizing the sim-to-real gap. Moreover, users will be able to test and integrate their own hardware kits within selected simulation environments in the Robora Dapp, ensuring seamless compatibility and safer real-world implementation.show more

𝗥𝗼𝗯𝗼𝘁𝘀 𝗱𝗼𝗻’𝘁 𝗻𝗲𝗲𝗱 𝗺𝗼𝗿𝗲 𝗱𝗲𝗺𝗼𝗻𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻𝘀. 𝗧𝗵𝗲𝘆 𝗻𝗲𝗲𝗱 𝘁𝗼 𝗹𝗲𝗮𝗿𝗻... 𝗳𝗿𝗼𝗺 𝗳𝗮𝗶𝗹𝘂𝗿𝗲 — 𝗮𝗳𝘁𝗲𝗿 𝘄𝗮𝘁𝗰𝗵𝗶𝗻𝗴 𝗵𝘂𝗺𝗮𝗻𝘀. Most robot learning systems assume failure is the end of learning. In our new work, we study whether robots can improve after deployment by learning from their own failures, without any human intervention, teleoperation, or corrective labels. The key idea is simple: human videos contain structure about how the world works. We use them to learn cross-embodiment representations of action, dynamics, and value, enabling a shared predictive space between human behavior and robot experience. This allows a new learning loop: 👉 pretrain on human videos 👉 deploy robot policy 👉 observe failures 👉 reinterpret failures using human priors 👉 improve autonomously We evaluate this across 7 real-world manipulation tasks, showing: 📈 40% → 81% success rate 🏆 Strong improvements over π0.6 RECAP and RISE ✔️ Zero human intervention during post-deployment improvement 🧬 Generalizes across robot embodiments and policy backbones A key finding is that explicit failure repair significantly outperforms failure reweighting, yielding substantially larger gains under identical data conditions (+25 pts vs +5 pts on the same π0.5 base policy). Overall, the results suggest a shift in how we think about robot learning: Human videos are not only for pretraining policies. They can provide the structure needed for continual self-improvement after deployment. 📄 Paper: 🌐 Project: I am grateful for working with the fantastic leads Hanzhi Chen and Anran Zhang, and our collaborators Simon Schaefer, Kejia Chen, Shi Chen, Daniel Cremers. Special thanks to Stefan Leutenegger for co-advising this project with me. ETH Zürich TU München Microsoft Check out Hanzhi's 🧵 for more detailsshow more

Disappointed with your ICLR paper being rejected? Ten years... ago today, Sergey and I finished training some of the first end-to-end neutral nets for robot control 🤖 We submitted the paper to RSS on January 23, 2015. It was rejected for being "incremental" and "unlikely to have much impact" Our resubmission to NeurIPS was also rejected It now has >4,000 citations (and more importantly, end-to-end training is widely accepted!) It's also cool to think about what's changed and what's the same -- - The network was 92k parameters and trained on ~15 minutes of data - The code was a combination of matlab, caffe, ROS, a custom CUDA kernel for speed, and a low-level 20 Hz controller in C++, all talking to each other. ROS+matlab was as bad as it sounds. - We pre-trained the encoder and did inference off-board on a workstation with a larger GPU. - We were paranoid about varying lighting messing up the network, so we did all the experiments after sunset (so long nights running experiments on the robot past 3 am) Now, we have manipulation policies that are far more dextrous, far more generalizable, and maybe on the cusp of breaking into the real world. :) (the paper:show more

🚀Thrilled to share what we’ve been building at TRI... over the past several months: our first Large Behavior Models (LBMs) are here! I’m proud to have been a core contributor to the multi-task policy learning and post-training efforts. At TRI, we’ve been researching how LBMs can help robots learn faster, better, and more efficiently. The key takeaways: ✅ We built an evaluation pipeline to benchmark LBM performance with real 𝐬𝐭𝐚𝐭𝐢𝐬𝐭𝐢𝐜𝐚𝐥 𝐜𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐜𝐞 ✅ Pre-training on hundreds of tasks makes models more robust—plus, we can teach new, complex tasks with 80% 𝐥𝐞𝐬𝐬 𝐝𝐚𝐭𝐚 ✅ The bigger and more diverse the pre-training, the better the results Check out our overview video, webpage and paper for more details: ✨ 🌎 📄 We hope this work helps move the field of robotics forward!show more

Introducing VL-JEPA: Vision-Language Joint Embedding Predictive Architecture for streaming,... live action recognition, retrieval, VQA, and classification tasks with better performance and higher efficiency than large VLMs. • VL-JEPA is the first non-generative model that can perform general-domain vision-language tasks in real-time, built on a joint embedding predictive architecture. • We demonstrate in controlled experiments that VL-JEPA, trained with latent space embedding prediction, outperforms VLMs that rely on data space token prediction. • We show that VL-JEPA delivers significant efficiency gains over VLMs for online video streaming applications, thanks to its non-autoregressive design and native support for selective decoding. • We highlight that our VL-JEPA model, with an unified model architecture, can effectively handle a wide range of classification, retrieval, and VQA tasks at the same time. by Delong Chen (陈德龙) Mustafa Shukor Théo Moutakanni Willy Jade Lei Yu Tejaswi Kasarla Allen Bolourchi Yann LeCun Pascale Fungshow more

A Letter to Our Community: The Road Ahead for... Robotics To our Community and Partners, As we step into 2026, our mission at Axis is clearer than ever: Constructing the definitive End-to-End Scaling Layer for Robotics. Our goal is to accelerate the transfer of diverse human intelligence into Robotics General Intelligence (RGI). By owning the critical path of intelligence creation, we are turning the physical limitations of robotics into a scalable, software-driven future. Here is our strategic outlook and roadmap for the year ahead. The Core Thesis: Simulation is the Only Way Out The path to RGI is currently blocked by Data Scarcity, Generalization Fragility, and Hardware Fragmentation. At Axis, we believe Simulation is the only way out. Our Simulation Data Platform and Data Augmentation Engine transform raw data into "Synthetic Gold". Backed by academic milestones like Roboverse, Skill Blending, and GraspVLA, we have proven that pure simulation can achieve the generalization required for the real world. We don’t just collect data; we architect it. The Engine: Why Crypto? We believe RGI should come from all, not a few. Crypto is not just a feature; it is the primitive that powers our entire ecosystem flywheel: - Incentive Mechanism: Democratizing contribution and rewarding the trainers and developers. - Assetization: Turning proprietary data and refined models into liquid, ownable assets. - Verifiable Workflow: We are opening the "Black Box" of AI. By bringing total transparency to the Task Generation → Data Collection → Model Training pipeline, we ensure every byte of intelligence is verifiable, traceable, and secure. 2026 Strategic Deliverables This year, we are committed to delivering three foundational pillars: - The World's Largest Training Dataset for Robots: A robot training set—diverse, high-quality interaction data at an unprecedented scale. - A Robotics Foundation Model: A universal robotic brain trained on our pure simulation and synthetic data, capable of robust cross-embodiment transfer and open-world adaptability. - Evolvable Robot Hardware: Robots deployed with Axis models that autonomously evolve through continuous interaction, turning every deployment into a self-improving node within our RGI network. The Ultimate Vision We are building more than models; we are architecting the Distributed Machine Economy. A future where every dataset, model, and robotic embodiment is a verifiable asset in a global, autonomous network. Thank you for building the future of intelligence with us✌️📷show more

This work makes a humanoid robot do simple parkour... moves by looking with a depth camera and choosing the right move on the fly. The big deal is that it turns lots of small human moves into long, real-time robot behavior, without hand-coding every transition or retraining for each new course. A humanoid robot is usually good at steady walking, but it often fails when it has to do fast moves like jumping up, vaulting, or rolling, and then keep going to the next obstacle. The hard part is that you cannot easily collect training data for every possible obstacle shape, distance, and mistake, so robots end up learning a few moves that only work in a narrow setup. This work starts from short clips of real human parkour moves, like stepping over, vaulting, climbing, and rolling. It uses motion matching, which is basically a smart “pick the next clip that fits best right now” search, to stitch those short clips into a long, smooth plan that looks like a human doing a whole course. Then it trains a controller with reinforcement learning (RL), which means the robot learns by trial and error to copy that plan while staying balanced and not falling. After training separate expert controllers for different moves, it compresses them into 1 controller that uses only onboard depth sensing and a simple “go this fast in this direction” command. In real tests on a Unitree G1 humanoid, it can clear multiple obstacles in a row, adapt when obstacles get moved, and climb a wall up to 1.25m.show more

Gemini-powered robot can now effectively debug itself! I've been... obsessed with two main questions in robotics: can robots learn from their own mistakes without humans in the loop, and how much can we leverage synthetic data? Spoiler: yes, and it's surprisingly elegant once you have the right primitives in place. The architecture is fairly simple (and optimized for GPU_Poor users): Component I: Gemini Brain ♊️ - Gemini 2.0 Flash analyzes all training episodes through both camera perspectives - Gemini 2.0 Pro creates a summary of training data, highlighting biases, limitations, etc. - Train policy p0 on this initial data, run evaluation episodes - Ask Gemini to categorize successes vs. failures (more insightful than you'd expect) - Based on both analyses, Gemini generates specific augmentation recommendations What's interesting here isn't that we're using LLMs for robotics - it's that we're closing the loop between perception, failure analysis, and targeted data generation. Component II: Data Generation with Scene Consistency The tricky part was maintaining consistency across both camera perspectives while generating new data. Three current augmentations: - Frame flipping and polarity reversals - Grounded-SAM + OpenCV for object color manipulation - Gemini to identify empty space and generate distractions in the scene …and repeat, ha! I'm using the so100 robot arm and Sarah’s Vintage from Hugging Face. And the APIs and models in Gemini family are Ace! Thank you Logan Kilpatrick Patrick Loeber and team for this. In thread The Circus of Making It Actually Work🧵:show more

Multi-robot learning is getting a serious boost! 📚 Researchers... have extended Isaac Lab to train heterogeneous multi-agent robotic policies at scale. The new framework supports high-resolution physics, GPU-accelerated simulation, and both homogeneous and heterogeneous agents working together on coordination tasks. They benchmarked different approaches (MAPPO: Multi-Agent Proximal Policy Optimization and HAPPO: Heterogeneous Agent PPO) across six challenging scenarios and showed that large-scale multi-robot training is not only feasible, but efficient. It’s an important step for real-world robotic collaboration, where teams of robots need to coordinate, split tasks, adapt roles, and interact dynamically, not just operate as identical clones. The code is open-source, and it pushes Isaac Lab closer to what robotics actually needs: scalable, physics-driven environments where many different robots can learn to work together. Here's the project page: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →show more

The fourth edition of Ather Community Day is here!... The idea behind these community days actually goes back to 2016. Back then, we were a startup without a product. We started Community Meetups as a way to tell our stories and, just as importantly, learn from our owners. Those conversations shaped not just how we talked about Ather, but what we built. As Ather grew, those meetups grew with us. In 2023, they became the first Ather Community Day with around 700 people. Last year, nearly 5,000 people joined us. While the scale changed, the purpose luckily hasn't. Community Day has always been where we share the thinking behind our products. The engineering decisions. The design trade offs. The details our teams spend months, and often years, obsessing over. Some of those stories might seem small. In 2024, we spent five minutes on stage talking about a storage box 🤯. Last year, we explained how we reinvented cruise control. Most product launches don't dwell on these details. Most product launches are not aimed at the customer at all. We do, because we believe they're what make the experience feel effortless. This year is a particularly important one for us. We'll unveil the first scooter built on our all new EL platform. Our biggest product launch yet. EL is our first all new vehicle platform since the original 450. We built it with a simple objective. Bring the Ather experience to a much larger audience without compromising on what makes an Ather, an Ather. It's our entry into the mass market scooter segment and the beginning of the next chapter for Ather. And EL is just one part of the story. We'll also share what's next across the Ather ecosystem and give everyone a closer look at some of the work happening inside our engineering and R&D teams. Look forward to meeting you all in Bengaluru on 29th August!show more

The term "continual learning" has become overloaded if you... see it as an ML problem. One classic thread is about memorization: regularization-based continual learning methods, such as EWC, MAS, and SI, estimate which parameters mattered for previous tasks and resist changing them too much. One modern thread is about adaptation: test-time training and inference-time learning methods, such as TTT, adapt part of the model on the incoming test stream before making predictions. These are sometimes discussed as separate threads. But in modern scalable architectures, I think they are better seen as complementary constraints: a model that learns quickly at test time also benefits from a mechanism for deciding what not to forget. In our #ECCV2026 paper, we study this in large-scale 4D reconstruction: how to build fast spatial memory that can adapt over long observation streams while reducing collapse and forgetting. Instead of using fully plastic test-time updates, we stabilize fast-weight adaptation with an elastic prior that balances adaptation and memory. Key ideas: - Elastic Test-Time Training: Fisher-weighted consolidation for fast-weight updates - EMA anchor weights that provide a moving reference for stability - Chunk-by-chunk inference for long 3D/4D observation streams We show that this scales across large 3D/4D pretraining settings, including both LRM-style and LVSM-style models, and improves reconstruction across benchmarks including Stereo4D, NVIDIA, and DL3DV-140. We release model checkpoints across different design choices: resolution, post-training curriculum, and whether the model uses an explicit 4DGS intermediate representation. - Homepage: - Paper: - Code: - Models: This work is co-led with Xueyang Yu, contributed by Haoyu Zhen Yuncong Yang, and advised by Michigan SLED Lab Chuang Gan.show more

I had a good meeting with U.S. Special Presidential... Envoy, General Keith Kellogg Keith Kellogg. Ukraine is deeply grateful to the United States for its support, and we value that President Trump President Donald J. Trump is so determined to achieve real peace. It is very important to realize all the results—political, defense, and economic—that were achieved during our meeting together with European leaders in Washington. Undoubtedly, it was a successful Summit, a demonstration of true unity between Europe and America. Ukraine, as always, is uniting the world. We value the United States’ readiness to be part of the security architecture for Ukraine, and our teams are actively working on shaping it. We expect that the key foundations of security will be defined shortly. We discussed how we can influence the Russians, compel them to engage in real negotiations, and end the war. Sanctions, tariffs—everything must remain on the agenda. We are ready to engage in a format of leaders. This is the format needed to resolve the key issues. Now, the same readiness is needed from Moscow. Military cooperation is important for both Ukraine and the United States, and there are two strong opportunities—an agreement on arms procurement and an agreement on drones that could significantly strengthen our arsenals. We are maintaining momentum in our work within PURL. This is an important instrument for procuring American equipment funded by partners, and we are now actively working on engaging additional countries. And, of course, the humanitarian track. The return of all abducted children. We strongly hope that America, President and the First Lady of the United States First Lady Melania Trump will continue to make personal efforts to bring back all children illegally abducted by Russia. Thank you!show more

This vandalism at Zimbabwe’s new Mbudzi Interchange is a... perfect example of what Dr Solomon Guramatunhu always reminds us — that Zimbabwean leaders are a reflection of Zimbabwean society and the Zimbabwean mindset. Our leaders do not fall from the sky; they come from our communities. What we are seeing here is no different from a leader who loots public funds. Public funds are meant for the public good. When ZANUPF loots national resources, it is not different from the Zimbabwean citizen who goes to an interchange and steals cables. Both acts are theft, both are sabotage of the common good, and both expose a destructive mindset that holds the whole nation back. It is exactly the same behaviour that South Africans have been complaining about us for years, when some Zimbabweans vandalise public infrastructure and steal cables across the Limpopo. We are quick to call that xenophobia, but what then do we call it when we are destroying our own country with the same reckless disregard? We all know that in Zimbabwean homes, from the poor to the affluent, there are lithium batteries stolen from mobile phone towers in South Africa and sold cheaply in Zimbabwe. When South Africans complain about this, we dismiss them as being xenophobic. Here is an example of us doing the same thing in our own country. We are destroying and stealing from ourselves. When the lights fail, the interchange will be plunged into darkness, and people will be mugged and killed because there is no lighting. This is wrong. We cannot build a better country with the same hands that destroy it. It can’t!!!!show more

Even after breaking into the house, stealing everything, and... brutally attacking @barbiekyagulanyi, as if that was not enough, they now want to force their way in again. Who knows what their real motive is? We are dealing with bandits. We are dealing with real terrorists in uniform. Our parents handed this nation to uncivilized pastoralist-minded war mongers, people trained only in bush warfare — where all they learned was to hide in the bush, kill people, quarrel with animals, and rule through fear instead of reason. No wonder they can harm anyone at any time without conscience. They don’t know how to exist within a just society. Look at how they have destroyed our country. Give them something beautiful, and when you return, it is never the same. Shame on them. May we find the wisdom, courage, and unity to rescue ourselves from these dark forces. 💔 #FreeUgandaNowshow more

It's 2030 and you are reviewing humanoid robots. A... Tesla. A Google. An Apple. An OpenAI. A Meta. A Figure. And a bunch of Chinese-made ones. Which one is best, and why? I think the Tesla understands the world much better. Why? There were eight Teslas around me on the freeway today. Start there. No other robot company has that data. But my robot is parked at the local high school twice a day. Its cameras see humans in all of our weirdness. How we move. Where we go. Where we walk. Who we talk with. What you are wearing. Whether your hair was combed this morning. That data will lead to robotics breakthroughs. Apple might keep up with its Vision Pro data, but it is too freaked out by the privacy implications of using said data. (On the front are six cameras and a couple of TOF -- Time Of Flight -- sensors that can see everything in your home in great detail). Google has a lot of data, for sure. All my: 1. Email. 2. Calendars. 3. Photos. 4. TV watching behavior. 5. Contacts. 6. Documents and spreadsheets. 7. Files. 8. Location data. So I expect Google's robot will be attractive to many. But how do you see the others shake out over the next five years? Make some guesses. But remember what an AI pioneer told me years ago about AI: it's all about the data. The Chinese ones have huge advantages: the Chinese have more data on their citizens, and many more citizens to boot AND they can make robots cheaper than we can. But now that you know OpenAI is building its own robot you have caught wind of what I've heard from many in San Francisco and Silicon Valley: that humanoid robots are the real prize of AI and will be highly profitable for those that can make them and find customers willing to buy them. Here, too, I learned long ago never to bet against Elon Musk. Will you?show more

Two best friends, both 22. They film their workouts,... AI edits everything, and brands pay them around $6,800 a month — combined. Here's how splitting the work doubled the income. They started the way most people do: training together anyway, filming a little for fun. The difference came from noticing what solo creators never can — two people in one frame is a different product entirely. Here's why brands pay more for a duo. A girl training alone is content. Two friends training together is a story — the spotting, the laughing between sets, the shared routine people wish they had. Viewers don't just want the workout; they want a gym friend. That feeling gets saved, shared, and tagged: "us." Tags are free distribution, and duo content collects them at a rate solo clips never touch. The production splits clean. One phone, propped low. They train like they always did — one on the treadmill while the other lifts, then swap. AI takes the raw footage and handles the rest: cuts, captions, on-screen text, scheduling across platforms. Neither of them has opened an editor. The channel runs on friendship plus a system. The money, split honestly. An activewear brand pays ~$2,800/month — two matching outfits in every clip is double the catalog space. A supplement brand adds ~$2,400 for the same reason: two credible users beat one ambassador. Creator payouts and affiliate links on the gear stack roughly $1,600 more. About $6,800 a month, divided two ways — $3,400 each, at 22, for training they were doing together regardless. Tools: ~$40, also split. The part worth stealing: everyone waits until they're "ready" to create alone. These two skipped that entirely. The friend you already train with is a co-founder you haven't asked yet. Full playbook below👇show more

FX1 is proud to announce our latest strategic partnership... with Muhdo, the pioneers of DNA and epigenetics-based performance tracking, to redefine how combat sports athletes optimize their training, recovery, and career longevity. At the heart of this partnership is MuhdoFightHub, an innovative platform that empowers fighters with real-time insights into their bodies, using genetic, biometric, and lifestyle data to enhance performance, reduce injury risk, and maximize potential. By integrating FX1’s AI-powered combat sports intelligence into Muhdo’s Fight Hub, we are building the ultimate fighter performance ecosystem—an all-in-one app that merges AI-driven fight analytics with cutting-edge health science. Beyond empowering athletes, Muhdo will also enhance FX1’s fighter profiles by providing athlete lifestyle data, adding a new and innovative dimension to FX1’s sports intelligence ecosystem. This level of deep, personalized data does not exist in mainstream sports data platforms, giving FX1 an exclusive advantage in how fight analytics are captured, analyzed, and commercialized. Together, FX1 and MuhdoFightHub are pioneering the future of combat sports intelligence, giving fighters the tools they need to compete at the highest level, backed by science and AI-driven insights. We’re excited to bring this game-changing innovation to the world of combat sports—this is just the beginning!show more