A robotics panel. NYC Tech Week. From Prototype to Production. 4 founders. 200 people. @standardbots, @Ultraroboticsco, @gen_intuition, @faunarobotics. - skip to 2:37 for what does "deployment" actually mean? - skip to 9:19 for what two weeks of perfect lab testing couldn't predict - skip to 15:50 for why China installs 9x more robots than the US per year - skip to 19:40 for the investor demo secret nobody talks about - skip to 32:40 for the honest conversation about teleop nobody wants to have - skip to 51:27 for why everyone selling robotics data is going to lose

Coding a robot with GPT 5.5! A 7dof robot arm w/ functional kinematics. [📍 bookmark, it’s open source] An open source harness for generating 3D models with your favorite coding agent. Custom gui, and STEP parts/assembly, 100% generated in Codex. Thanks for sharing, @soft_servo!!! 📍 lnkd.in/df-F2Ziq ——- Weekly robotics and AI insights. Subscribe free: 22astronauts.com

Dear algorithm: please show this to people interested in robotics, embodied AI, automation, and manufacturing. I’ll be at MACHINA in Paris tomorrow and Tuesday. Thank you.

Boost real-robot policy success & speed using only cheap simulation, without needing any extra real data: - 2.4× success 37% faster execution - Half-day iteration with new retry behaviors Real-robot imitation policies for dexterous manipulationby learning to steer diffusion or flow models in simulation without collecting extra real-world data. The approach freezes the base policy trained on real data and uses RL only to select successful latent inputs within the policy’s support, avoiding sim-reality gaps from contact dynamics that plague full finetuning. Results show 2.4x higher real-world success rates and 37% faster execution across tasks like block picking, with quick half-day iteration from setup to deployment while enabling retry behaviors absent in base policies. Thanks for sharing, Abhishek Gupta / @abhishekunique7! 📌 weirdlabuw.github.io/score/ —— Weekly robotics and AI insights. Subscribe free: 22astronauts.com

Do you even lift? A teleoperation framework that enables full-size humanoids to handle heavy payloads: Like this 175cm, 65kg L7 robot that handles up to 24kg using raw, noisy consumer VR inputs for human intent tracking. Lab demos show the L7 robot walking stably while lifting two 12kg kettlebells or a 10kg water bottle, alongside household tasks like carrying grocery baskets, interacting with a fridge, and serving items, highlighting real-world versatility. The work also validates HEFT on the Unitree G1 platform for noisy VR teleop and high-dynamic whole-body tracking, with open training code, checkpoints, and a project site at heft.axell.top. Thank you so much for sharing, @Axell_wppr! 📌 Website & more demos: heft.axell.top G1 & L7 training code/checkpoints: github.com/Axellwppr/motion_… ——- Weekly robotics and AI insights. Subscribe free: 22astronauts.com

🤣🤣🤣

i’m shocked that MrBeast lost $120m on his youtube business, which is now essentially lead gen for his chocolate bars (among other products), which have a 2.5 rating at target. meanwhile, every american is being bombarded by chinese ai shorts, ads, and mini dramas that cost $5 to generate. content, models, commerce, etc. are all getting temu-fied. watch the cost!!!

things every serious robotics team ends up rebuilding from scratch: – sim → logs → eval loop for policies (isaac, mujoco, custom glue) – fleet management that isn’t just ssh spreadsheets – observability for ros2 / can / fieldbuses that ops can actually use – a data layer tying missions, bags, labels, and models together – safety cases procedures that pass real audits, not just internal vibes[automate] – operator tools so non‑engineers can debug, teleop, and recover safely most of the “robotics stack” isn’t off‑the‑shelf yet. every team is quietly rebuilding the same six systems.

so based on this tweet there's probably >100 startups working on VLM for robotics and ego data labelling if you work on this, prove that your solution achieves >99% accuracy on a complex task at all conditions. if I can't trust the labels I'm not going to train on the data

If you work on flexible/soft object manipulation... this one is for you: Teach robots new skills directly on real hardware. {No weeks of simulation or endless trials needed} Enabling a robot to tie a dynamic flying overhand knot in the air using one human demonstration and under 10 hardware trials. The approach employs Task-Level Iterative Learning Control that focuses updates on the critical rope collision point via a learned inverse model, allowing rapid adaptation directly on physical hardware without extensive simulation or data. The robot achieves 100% (!!!) success rate post-learning across varied ropes like chains and tubes, plus different human demo styles, with strong transfer in 2-5 trials. Thank you for sharing, @_krishnasuresh!!! 📌 Website: flying-knots.github.io/ Video: youtube.com/watch?v=FLiILOyQ… ——- Weekly robotics and AI insights. Subscribe free: 22astronauts.com

I'm increasingly interested in the problem of *multisensory continual learning*, since it feels inevitable for robotics. Unlike vision, many robot sensors (e.g., force/torque, tactile, audio) are highly task- and system- specific. It's unrealistic to expect a single pretraining dataset to contain every future sensor. And as robotics evolves, we'll keep building new sensors. So the question is: Can we plug a new sensor into a pretrained vision-only foundation model without forgetting everything it already knows? Better yet, can the new sensor actually improve the model's existing vision-based skills? That's exactly the question that motivated MuSe 👇

We’ve been approaching reward supervision for robots the wrong way. I think freeform preferences are part of the answer. A short 🧵

It's so satisfying to watch the natural, dexterous behavior that our Play2Perfect policy produces. Play2Perfect first learns a base policy with generic skills like grasping, in-hand re-orientation and goal reaching. Then it finetunes this policy for a specific assembly task.

This week, during Davos Tech Summit you can see a robotic startup that is tackling Europe’s old infrastructure… Critical infrastructure. From old hydro networks to the world’s shipyards. 60% of it hasn’t been touched in 40 years!!! No way this will not be done by robots! I’ve been working closely with the Nio Robotics team lately, and what they’ve built is wild... Aru. Built as a platform. To do welding in ships, to decoat on a pendstock, painting, inspection, maintenance, and many more jobs. In confined, hazardous spaces humans can’t (or shouldn’t) work in. AND: they are building the software stack that’s making Aru into a full autonomous operator! But I am biased. Want to see it live? It’s part of Davos Robot City, running through July 4 — go see it in person. More here: nio-robotics.com

Come see the @gen_intuition MTS and the MOPS (member of ops staff) at ICML! @micheli_vincent @EloiAlonso1 @hu_anth @JSwingos and many more!

Davos becomes Europe's first Robot City. Starting today, the Davos Tech Summit turns the entire town into a live testing ground for physical AI: Humanoids, quadrupeds, drones, and other systems working in real-world settings across the city, not on a stage. Speakers include Roland Siegwart (ETH Zurich), Marco Pavone (NVIDIA), Ryan Gariepy (Rockwell), and 80 other experts. Co-initiated by @ETH_AI_Center, with co-initiator Alex Ilic describing it as "the bridge between the digital intelligence of today and the physical intelligence of tomorrow." Deployments include @UnitreeRobotics, @anybotics, @AGIBOTofficial, Pudu Robotics, Starship Technologies, and Nio Robotics, among 35 teams. I'll be there as a jury member for the Robot City Award: judging systems in the field, not in a pitch deck. The question that matters most in robotics right now: not "can it walk," but "can it work." If you're in Davos this week, let's connect. 👋 More: davostechsummit.com ——- Weekly robotics and AI insights. Subscribe free: 22astronauts.com

This model helps robots learn tricky moves and tasks much faster and better. How? By turning their actions into a short, smart code! 📌 { Open-source code, arXiv paper, and project page } Ahad J. introduces Neural Action Codec (NAC), which models robot actions as audio waveforms to create a highly compressed vocabulary for training vision-language-action policies. NAC achieves 27x token space compression, near real-time decoding, and better performance by adapting audio techniques while removing mel spectrogram losses and emphasizing adversarial losses. Policy comparison videos demonstrate NAC outperforming baselines like Diffusion, VQVLA, and OAT in manipulation tasks. Open-source code, arXiv paper, and project page shared for replication: Thank you for sharing, @ahadj0! 📌 paper: arxiv.org/abs/2606.21372 code: github.com/ahadjawaid/nac project page: ahadjawaid.com/nac ——- Weekly robotics and AI insights. Subscribe free: 22astronauts.com

If you are interested in robotics, and you don't follow and watch @TheHumanoidHub for his robotics insights (here in China)... no one can help you.

A new framework for robotics to handle complex and contract-rich tasks! Delivers both high precision AND strong generalization/robustness. What kind of tasks? Contact-rich tasks like motherboard assembly, now being enabled by composing three simple behaviors: Semantic, predictive, and reactive. How? Via a shared SE(3) interface rather than using monolithic policies or rigid pipelines. The system integrates foundation models for scene understanding, video world models for trajectory prediction, and high-frequency tactile feedback for real-time corrections. Achieving 96.7% success (!!!) on precision assembly tasks and demonstrating generalization to everyday manipulations without task-specific retraining. The attached video shows a robot successfully inserting CPU, RAM, and GPU components while recovering from human-induced perturbations, highlighting CoStream's robustness and modularity in real-world settings. Thank you for sharing, @HaonanChen_ !!! 📌 Source: costream-simple.github.io Paper: arxiv.org/abs/2606.26423 ——- Weekly robotics and AI insights. Subscribe free: 22astronauts.com