RECORDED IN ISAAC SIM
01
Magnetic field
A UR5e with a SCHUNK EMH magnetic gripper picks and places a steel plate, with field lines and holding force computed on the GPU.
Building the missing physics layer of robotic simulation.
Bridging the gap between modern Physical AI models and traditional physics-based modeling, for real autonomous deployment.
Beyond mechanics
Robotics simulators are built for mechanics: rigid and deformable bodies, joints, contacts, cables. But real robots don’t fail on mechanics alone. They fail on the physics underneath: thermal, electrical, electrochemical, electromagnetic.
We bring that physics into the simulation stack. For example:
- Actuator Joule losses and heat transfer
- Battery state of charge and health (SOC/SOH)
- Magnetic fields and flow dynamics
- Space robotics: free-floating dynamics and radiative cooling of actuators
- Tactile, piezoelectric and thermal sensing
Our modules plug into NVIDIA Isaac Sim and the Newton physics engine, and run on the GPU. You keep the scale of reinforcement learning, synthetic data and world models, and close the sim2real gap.
Use cases
RECORDED IN ISAAC SIM
02
Fluid dynamics
The UFACTORY Vacuum Gripper on a Lite 6 picks a 20 g plastic film, simulated as deformable cloth in Newton. Vacuum build-up, air leaks and seal loss come from one air balance, computed with Warp.
RECORDED IN ISAAC SIM
03
Actuator Joule heating
A Unitree G1 walks through a warehouse carrying a box, driven by Unitree’s own locomotion policy. Every joint torque is the effort sent to the physics engine, and each actuator’s heat loss is computed from its torque and speed on the fly. Joints and the panel are coloured by that loss, so you see which motors are working hardest before they overheat, the first step of a motor thermal management system.
RECORDED IN ISAAC SIM
04
Microgravity and radiative cooling
Example of the Canadarm, the Space Shuttle’s robotic arm first flown in orbit in 1981: 6 DOF, accurate joint limits and mesh from NASA. As orbital datacenters grow, robotic maintenance in space will be required, and radiative cooling of actuators combined with microgravity conditions will need to be simulated.
The missing physics that is impacting the Sim2real workflow
One example among many: the same vacuum pick of a thin plastic film, run in Isaac Sim with Newton and on the real Lite 6. In simulation the suction cup grabs the film. On the real table the film stays flat.
Two forces that Newton and Isaac Sim do not simulate natively could explain the change in behavior:
- Electrostatic force. The fake-wood table acts as the dielectric and the film starts out charged. That induces an attractive force toward the table many times the weight of the film.
- Capillary attraction. If the thin film is slightly wet (humidity, previous handling), capillary adhesion holds it to the surface.
Optimized computing
GPU-accelerated simulation
Running on NVIDIA Warp, a purpose-built open-source Python framework that delivers GPU acceleration for computational physics, AI and optimization workflows.
AI-accelerated physics
Combining CFD and FEA solvers with machine learning to make traditional physics simulation much faster. Neural-network surrogates for traditional solvers (OpenFOAM, Ansys, STAR-CCM+).
Full deployment simulation
Most simulations train a policy on short episodes in clean, isolated environments. We simulate the full deployment: long, autonomous operation, with the physics that decides whether it holds and the metrics that follow: temperatures, state of charge, holding forces, time before a forced stop.
Accurate physics lets you trust the simulation where real tests are too long, too remote or too risky. For VLAs, which map pixels straight to actuator commands, physical limits are the safety guardrail.
A policy ready for real deployment
A PPO policy trained in Isaac Lab learns what its reward terms measure: velocity tracking, stability, smoothness. Nothing tells it about winding temperature, state of charge or a thermal cutoff.
Our modules expose those quantities as per-actuator tensors on the same GPU. Add them to the observation space and as reward terms, domain-randomise the thermal constants, and the same PPO run learns the deployment constraints before it meets the real robot.
g1_thermal_rewards.py
Isaac Lab · Unitree G1