State Labs

Came for the chips, stayed for Jensen's new jacket. 🧥 Spent the day at #GTC26 seeing the next gen of humanoids. Atoms are the new bits. #PhysicalAI #NVIDIA #EmbodiedAI

☕️ Expect high-density discussions on: 🔸 Scaling AI infrastructure & optimizing compute 🔸 Moving from frontier models to production-grade apps 🔸 Bridging global AI innovation with the Asian market We’re bringing together the minds shaping the next gen of AI. 🧠

Are you just following AI — or building with it? 🛠️ We are hosting a curated gathering during #GTC26 for builders focused on what actually works in production. Frontier AI meets real-world scale. RSVP 👇 luma.com/j5ew4n21 @googlecloud @nvidia

Day1 at #EmbeddedWorld 2026! 👋 You can really feel how fast edge AI is moving👀come meet us! #ew26 #PhysicalAI #Robotics #EdgeAI #OnDeviceAI

@after_ephemera Good question. We haven't published yet but i can tell you it is fast (>30Hz).

@STL_AI What kind of inference frequency were you able to achieve?

SFT can teach robots to mimic. It doesn’t teach them to recover. We trained a VLA SFT baseline (pi 0.5) on ~270MB (30 demos/scenario). Imitation looked great, but small shifts dropped success to ~30%. Full write-up: statelabs.ai/blog/1

@physical_int impressive

We’ve developed a memory system for our models that provides both short-term visual memory and long-term semantic memory. Our approach allows us to train robots to perform long and complex tasks, like cleaning up a kitchen or preparing a grilled cheese sandwich from scratch 👇

full article 👉 statelabs.ai/blog/4

SFT can make robots look competent, until conditions shift. In our test, a minor change dropped success to 29.2%. With RL: 74.5% (and generalization stops collapsing). Quick summary below. Full notes in reply 👇

Full write-up (ManiSkill + PPO + rollout/distillation loop): statelabs.ai/blog/2

SFT makes robots look competent—until the task shifts. In our benchmark: 29.2% → 74.5% after VLA-RL. Stress tests (never trained): • distance ×3 → 53.7% • appearance shifts → 43.3% What breaks real-world robotics more: generalization, latency, or cost?

AI belongs on the factory floor. Updates soon as we move forward.

We are bringing secure, distributed AI from the cloud to the concrete. By integrating OpenTMP CIS with Zekeep’s robotic platforms, we’re solving the hardest problems in industrial automation: ✅ Sub-millisecond Latency ✅ Safety-First Engineering ✅ Enterprise-Grade Security

Also tackling deployment: pi-scale models are heavy for edge; cloud inference adds latency. We’re pushing generalization + efficient edge running.

Next: Reinforcement Learning to close the 70% gap, recovery, robustness, and real-world generalization.

Pushing forward on dexterous manipulation with the Linker Hand from Linkerbot.👋 Seeing learning-based control run on high-DoF hardware in the real world has been really encouraging. Early signals look promising, and there’s a lot of potential ahead.

ROMI Lab is a true powerhouse of young, brilliant talents in robotics control. It’s refreshing to work with a team that is so grounded, energetic, and focused on making things work in the real world.💪

Strategic collaboration with ROMI lab (a robotics research lab under The Hong Kong Polytechnic University)！ Both parties agree to jointly research on VLA + Robotics and will disclose more details under this framework as progress is made. Let’s build the future of embodied AI!

@physical_int Impressive

We got our robots to wash pans, clean windows, make peanut butter sandwiches, and more! Fine-tuning our latest model enables all of these tasks, and this has interesting implications for robotics, Moravec's paradox, and the future of large models in embodied AI. More below!

Thrilled to be awarded by the Shenzhen Robotics Association! Onwards and upwards, let's keep pushing the boundaries of what's possible! 🚀🤖 #Robotics #Award #Technology #FutureTech