Runhan Huang @RunhanH

wow!! GPT-6 Astra can rapidly adapt to a new robot embodiment and a new task 🤯🤖 We asked it to put a cake into an air fryer in the benchmark we’re building, using simple tool APIs including CuRobo-based motion planning. Its first attempts were chaotic. After repeated interaction, it accumulated environment-specific experience and learned to reliably complete the task end to end. 🍰➡️🔥 This is generalization far beyond VLAs and WAMs.

Remember the car task in the DARPA Robotics Challenge? I asked Astra to tackle getting a humanoid into a Tesla—in Mujoco. Two videos: learning through my feedback, then reusing that experience for direct execution using onboard egocentric perception and in-context adaptation.

We put GPT-6 Astra in the RoboDojo. 🥋🤖 The RoboDojo Team conducted a comprehensive evaluation of GPT-6 Astra as an embodied agent, including: • RoboDojo Sim & Real, compared with GPT-5.5 and DeepSeek-Flash • Humanoid high-level control • Dexterous piano playing with RoboPianist 🎹 • A systematic study of in-context learning (ICL) Our key takeaway: GPT-6 Astra demonstrates remarkably strong semantic and spatial understanding, together with impressive in-context adaptation. At the same time, physical commonsense remains a clear bottleneck — revealing an important gap between understanding the world and truly reasoning about its physics. Full report & demos: robodojo-benchmark.com/report/gpt-6-a… @_wenbozhang (project lead), @wenhaocha1, @frankzydou, @JinWeiyang18434, @YutaoOuyang, @minifullcapsule, @x_h_ucb, @YutaoOuyang, @YueChen614

Introducing Hypit: Clone any viral video with AI agents. 1 clone, 100 variants, 100M views. Hypit lets your AI agent (Claude Code, Codex...) clone any viral video. Paste any viral video link from TikTok, Instagram, or YouTube into your agent. Hypit clones it into a complete agentic video workflow: footage, captions, B-roll, effects. GitHub: github.com/hypit-ai/hypit Key Point: - Arcads: $220 / month - Higgsfield: $129 / month - Creatify: $99 / month - Hypit: FREE 🌟 Build the video creation harness for AI Agents. Redefine how videos are made. Like + comment "Hypit" and I' ll send you 100 free AI avatars, each with its own unique voice. And if you like what we’re building, give Hypit a 🌟 on GitHub.

How can we improve robot generalization beyond collecting more demonstrations? A robot can use inference to work out how to solve tasks it was never explicitly trained to perform. With learned models of the world, we can plan our future actions and goals before acting. I wrote a research perspective, Generalization by Construction, on how learning can be combined with inference for flexible generalization. yilundu.github.io/perspective/ge…

🤖 World Action Models predict the future to act better. However, WAM research remains fragmented in code, ad hoc in design, and unprincipled at scale. 💡 Today, we release 𝗢𝗽𝗲𝗻𝗪𝗔𝗠: 𝗔𝗻 𝗢𝗽𝗲𝗻, 𝗠𝗼𝗱𝘂𝗹𝗮𝗿 𝗘𝘅𝗽𝗹𝗼𝗿𝗮𝘁𝗶𝗼𝗻 𝗧𝗼𝘄𝗮𝗿𝗱𝘀 𝗦𝘆𝘀𝘁𝗲𝗺𝗮𝘁𝗶𝗰 𝗪𝗼𝗿𝗹𝗱–𝗔𝗰𝘁𝗶𝗼𝗻 𝗠𝗼𝗱𝗲𝗹 𝗣𝗿𝗲𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 to change this. 🌐 openwam-official.github.io 🧵1/14

Fun experiment: what if frontier models were x OOMs faster, y OOMs cheaper, and fit on an edge device? I had GPT-6 Astra control a simulated Go1: 50 Hz joint targets, like an RL policy. Physics paused between calls. 250 inferences → 5 seconds of walking.

Jitendra MALIK @JitendraMalikCV

2 weeks ago

@phillip_isola I disagree. You need high control frequency controllers/policies that deal with forces and contacts. If you believe what you said, here is a simple challenge: implement policies for quadruped locomotion e.g.(vision-locomotion.github.io RSS 2021, CoRL 2022) in your favorite LLM and

3 15 297 62K 112

Here is a follow up. GPT-6 Astra didn't write code to implement a trot gait and use those numbers. But it could well have memorized the raw joint angles of a trot gait and not actually use the robot state feedback. So here is a Go1 climbing stairs. (This is blind locomotion and again physics is paused between inference calls) I have attached my two prompts as well. Back even in 2023, previous works exploit the impressive in-context learning capabilities of GPT-4 and show trajectory level control by seeding with expert trajectories both for locomotion and manipulation. Here are two works: robot-learning.uk/language-model… prompt2walk.github.io Lastly, I have spent/wasted a whole year way back in 2024 with sim2real RL for custom quadruped robots with high gear ratios (because the motors were weak) and non-negligible leg mass. Now in 2026, it is relatively easy to get a RL policy working in the real world on a Unitree Go1 or G1 only because of the mountain of work done by the community on these platforms. Hardware is really hard. So please take the post in the spirit it was written in, the assumption was if inference latency, compute and cost was not a bottleneck.

keyboard_arrow_left Previous keyboard_arrow_right Next

Srinivas K @sri299792458

2 weeks ago

Fun experiment: what if frontier models were x OOMs faster, y OOMs cheaper, and fit on an edge device? I had GPT-6 Astra control a simulated Go1: 50 Hz joint targets, like an RL policy. Physics paused between calls. 250 inferences → 5 seconds of walking.

14 12 133 48K 75

I wanted to find whether GPT-6 hits its limit on robot dexterity. So I just gave it two videos, without states or actions, and told it to do real2sim and physical retarget to Wuji hands. Here are the successful physical rollouts — a Rubik's cube, then an egocentric human video.

Excited to share our work accepted to CoRL 2026 @corl_conf ! We present SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer — a unified policy that enables humanoid robots to learn and compose diverse soccer skills, achieving smooth skill transitions and long-horizon multi-skill behaviors. Project page: [SkillX project page](yzc0731.github.io/SkillX/) Paper: [SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer (arXiv)](arxiv.org/abs/2609.06718) Video below 👇 #CoRL2026 #RobotLearning #Robotics #AI #EmbodiedAI #HumanoidRobots #Soccer

Behavioral cloning mystery seohong.me/blog/behaviora… I wrote a new blog post about "mysteries" in behavioral cloning that appear with real-world robot data (e.g., overfitting is "good"). I also tried to demystify them and shared my thoughts!

Excited to share SLAMFormer-∞: an Infinite SLAM Transformer for unbounded frontend and backend processing. From monocular RGB alone, it jointly estimates camera trajectories + dense geometry—and scales to a 17 km urban drive. 🧵

We live in a multimodal world. We see, talk, act, and dream. Yet most LLMs still start with language pretraining. Why not train them natively with multimodal I/O from scratch? Because it’s SUPER HARD, adding modalities triggers training instability, design complexity, and often modal competition So what’s the path forward? Introducing: Towards Physics of Multimodal Pretraining (junlinhan.github.io/projects/physi…) We unpack the underlying mechanics of multimodal pretraining across 4 aspects: Knowledge Flow, Modality Synergy, Early Unification, and Recipe.

Action chunking is a mysteriously effective method. Modern large-scale imitation learning basically doesn't work without it. But why does it actually help? In our new paper, we try to break down the reasons. As the saying goes, what happened next might surprise you...

Andrew Wagenmaker @ajwagenmaker

2 months ago

Action chunking is a critical component in virtually all modern approaches to imitation learning for robotics. But why is it so critical, and do we really need action chunking? Check out our latest work to find out! (1/n) action-chunking.github.io

13 60 441 109K 429

Action to action flow matching is a new form of IL, bring faster inference speed. The key question is how much noise do we need in flow matching for multi modality. Human don’t need multimodal decision all the time, e.g., high multimodality when making choice and no multimodality when walking the same way. Then we have MARS policy from A2A, adaptively controlling the noise level. jingliangli.com/MARS_Policy/ Thanks for the podcast by @micoolcho and @chris_j_paxton

RoboPapers @RoboPapers

2 months ago

Diffusion Policy was one of the big breakthroughs that has enabled an explosion in real-world robot learning. However, it’s always had a weakness, which is that it works by computing a final action trajectory from random noise, which leads to high latency when predicting a final

1 12 80 29K 65

100K Hours = 4,166.67 days = 11.4 years real-world trajectories 57% success rate on the RoboCasa365 benchmark Maybe we're missing something

Xiaomi Tech @XiaomiTech_

2 months ago

Xiaomi-Robotics-1 is now open source!🎉 Xiaomi-Robotics-1 was pre-trained on more than 100,000 hours of UMI data and post-trained with over 10,000 hours of cross-embodiment data. This open-source release covers the complete pipeline from real-robot post-training to model

53 207 2K 133K 657

Another belated update: the recording of our RSS 2026 Workshop is now available! Check it out! Huge thanks again to all attendees, speakers, and co-organizers! youtu.be/fN9L5T60I3c

Haozhi Qi @HaozhiQ

4 months ago

🤖 Excited to announce the 4th Workshop on Dex Manipulation. How time flies! Join us and submit your work to our workshop! Speakers, schedule, call for papers, and more details are available: shorturl.at/APhXq Deadline: June 22, 2026 (AoE) See you in Sydney! #RSS2026

1 15 93 22K 24