PHYSICS_ENGINE_ONLINE
SYS_REF_SIM_GEN3

Parallel_Sim

gpu-native rigid body dynamics and contact solver. engineered for massive throughput, eliminating the cpu bottleneck in robot learning.

// OVERRIDE: MANUAL_CONTROL_LOCK

01 // Hyper-Parallelism

4,096+
Environments / H100
1.2M
Steps Per Second
~14MB
VRAM Per Env

by maintaining the entire simulation state on device memory, we bypass pcie transfer latency. the forward simulation, collision detection, and policy inference all execute within the same cuda context.

02 // Kinematic Solver

Mass-Matrix Solver

utilizes a custom articulated body algorithm (aba) optimized for warp-level execution on nvidia architectures. resolves complex kinematic chains without the drift associated with maximal coordinate solvers.

Contact Constraints

resolves rigid body collisions via a projected gauss-seidel (pgs) iterative solver. uses sub-stepping (typically 4-8 steps per control cycle) to prevent tunneling and ensure stable resting contacts for complex foot morphologies.

03 // Sim-to-Real Delta

Zero-Shot Transfer Metrics
Joint Angle ErrorACTUAL: 1.2°
Actuator Torque DeltaACTUAL: 4.5%
CoM Position DriftACTUAL: 2.1cm
* metrics collected over 100km of physical hardware traversal compared against deterministic replay logs. achieved via extensive system identification and domain randomization parameters.