Physical AI is one of the terms getting the most attention in robotics this year, and one of the least defined. Depending on who’s using it, it can mean humanoid demos, foundation models trained on physical environments, or just robots with more autonomy than before.
Most of our work in this space falls into two areas: simulation accurate enough to trust, and getting a trained policy from simulation onto a real robot through ROS 2 without rebuilding the pipeline each time. That distinction matters because most of the public attention goes to the model side, bigger world models, better foundation policies. The part that decides whether a robot actually ships is less visible: whether the simulation a policy trained in resembles the real environment closely enough, and whether the software running on the robot can execute what came out of that pipeline in real time.
Some examples from the last few months.
Simulation to Real Robot
We trained a simulated Boston Dynamics Spot to walk, navigate waypoints, and avoid obstacles from scratch, using reinforcement learning in NVIDIA Isaac Lab. No hand-tuned gait or motion capture reference, just RL learning locomotion and obstacle avoidance in sim. Read the post →
We applied the same approach to manipulators: RL policies trained in MuJoCo, deployed on real hardware through ROS 2, with the goal of making that handoff repeatable rather than a one-off demo. Read the post →
Modeling for Simulation
Getting a simulation to match reality starts with getting the model right. We wrote about modeling the 3-omniwheel base of LeKiwi, an open-source robot built on Hugging Face’s LeRobot project, in MuJoCo. Read the post →
Simulation Fidelity
We tested NVIDIA’s NuRec to see where it helps with visual fidelity in 3D robotics simulation, and where it doesn’t. Read the post →
We also looked at photorealistic synthetic data generation as a way to feed ML training pipelines with data that’s expensive or impractical to collect in the real world. Read the post →
From Lab to Field
Humanoids and AMRs are scaling quickly, but most stay at the demo stage longer than expected. We wrote about why: usually it’s the software stack, not the hardware, that isn’t ready for the field. Read the post →
Where This Fits
Our work here sits in simulation and ROS 2 deployment, not foundation model training. Teams that reach out to us at this stage are usually past the point where a demo is impressive enough. They’re trying to get a policy that works in simulation to hold up on real hardware, reliably, without months of custom integration work.
If that’s where your team is at, this is the kind of problem we work on.