🚀A belated post, but - such incredible two weeks for Dyna Robotics - From #Actuate by Foxglove in SF to Conference on Robot Learning (CoRL) + Humanoid 2025 in Seoul - and we couldn't have asked for a better run.
🤖 What we showed (boldly, and why it mattered):
1. Zero-shot, non-stop live folding demos in two totally new environments, 99%+ success rate.
2. At Actuate, our cofounder & chief scientist, Jason Ma delivered a keynote while our robot calmly folded for ~20 minutes on stage. If you’ve ever demoed under harsh stage lights, you know how unforgiving that is. 😅
3. In Seoul, we went even spicier: first time running our latest model on new hardware, out of box. Then we ran 3 days × ~7 hours/day, >99% success rate while researchers stress-tested us with random interventions and zero-shot T-shirts.
Each time we finished under interference, the room broke into applause. 🤖🙌
💙 What did people say?
"Impressive".
"A must-see demo on CoRL".
"I finally saw intelligence in robots."
“Hard to believe a ~30-person team built this in under a year.”
🧠 My takes on CoRL 2025 vs. last year:
1. Much bigger crowd and a clear shift in themes: from last year’s sim/manipulation focus to learning from human data, VLA/agentic stacks.
2. Hardware jump: lots of dexterous hands, production-grade teleop, tactile sensors, UMIs, plus the RDT2 drop (https://lnkd.in/gQ4fzGKF
). Expect policies trained on fresh hardware data hitting prod soon.
3. Near-term reality: debates aside on “how fundamental” VLA is to achieve physical AGI, scaling is still winning—the academia ↔ industry loop now feels very LLM-era.
4. Generalist wave: different bets/priorities on generalization (task, reasoning, embodiment). Dyna is zeroed in on high-performance task generalization.
🦾 Favorite demos I tried: Gemini Robotics by Google DeepMind (I donated data from my bracelets), and Unitree Robotics (my first time controlling a robot to🥊 and my hands shook during the flying kicks).
🤝 Community highlights:
1. Both our Actuate and CoRL happy hours were packed - over 120 brilliant researchers, engineers, and founders swapping notes late into the night.
2. We learned a lot from peers from China. The hardware iteration velocity is wild, supercharged by advanced manufacturing capacity.
3. Shoutout to Samsung Next for an excellent founder dinner - the BBQ was 🔥
🦾 Huge thanks to everyone who came through, and to the amazing team at Dyna Robotics for making the mission impossible happend Jason Ma Keith Williams Lindon Gao Claire Gauthier Justin Lazaro
🌍What's next:
If we didn’t get a chance to connect at the events, if you want to be invited for the next gathering, if you are interested in joining Dyna Robotics, my DMs are open.
#robots #embodiedAI #reinforcementlearning #agents #CoRL2025 #Humanoids2025 #Actuate #robotlearning #AGI
