Simulation is a core part of how we build and evaluate robots.
I spoke about simulation, robotics, and the role of interactive worlds at AI Engineer Singapore this past weekend!
Full talk : https://t.co/TNaa5Z0JUx
Huge shoutout to @SherryYanJiang@unprofeshme@agrimsingh@swyx and the whole organizing team! Definitely the most fun I've had at a conference. Would love to be back next year :)
Announcing HUD's RL environments for RSI hackathon! 🎉
Join us June 20–21 in SF if you're interested in RL and want to push the frontier forward!
(w/$100,000+ in prizes and compute credits 👀)
Directionally, I agree but I wouldn't phrase it as a clean tradeoff. Example: opening drawers. If the sim varies handle positions, drawer size, friction, lighting, etc., better generalization will help. But if the drawer isn’t actually articulated or the handle contact physics isn’t modeled correctly, the policy never learns how to open a drawer in the real world
They definitely overlap but aren't the same. Generalization is about whether a policy can handle new situations. Sim2real is about whether the simulated distribution actually covers the real-world distribution, like the physics properties, sensors, contacts, and materials, etc. Generalization helps, but it doesn't solve sim2real if the simulator is missing these variables
The internet gave language models their data for free.
Robots don’t have that. Every trajectory has to be earned through hardware, time, teleoperators, and real consequences.
Shrey’s piece on simulating the physical world for robotics is the clearest explanation I’ve read of why that data problem is so hard.
Must read if you’re serious about where robotics is going.
@tdkardum SimReady is an NVIDIA standard for 3D assets: https://t.co/DzHXmqozWi
Our platform automates scene authoring (articulated asset creation and contextually-correct placement in the scene) which you can run in Isaac sim, MuJoCo, etc
We're launching early access to Gizmo, our automated sim creation tool. From text and/or image inputs, our agent generates SimReady assets and scenes from dimensioned primitives, with correct affordances and articulation.
We're launching early access to Gizmo, our automated sim creation tool. From text and/or image inputs, our agent generates SimReady assets and scenes from dimensioned primitives, with correct affordances and articulation.
@aryanmadhaverma Not limited to the assets on the landing page! You can generate anything but some might need slight manual edits. Can you DM me your email? We’ll give you access
@xm_build still early but we track where the agent fails and users have to manually edit, and use that to improve our system over time (better prompting, tooling, and eventually training)
Sims are a crucial piece of the physical AI stack, and the hardest to scale. Check out our agent that solves the first step of that. Reach out for access!