Velocore
TECHNICAL SPECIFICATION WHITE PAPER

Sub-Millisecond
Physical Foundation Intelligence

How Velocore synthesizes continuous geometric deep learning, synthetic physics scaling, and bare-metal edge acceleration to achieve deterministic 1,000 Hz robotic control.

MODULE 01

SE(3) Equivariant Diffusion Policy

Classical deep learning treats 3D spatial rotations as separate unstructured coordinate vectors, requiring millions of redundant camera angles. Velocore incorporates the Special Euclidean Group SE(3) directly into the attention kernels.

When an object rotates in physical space, our internal tensor representations rotate equivariantly. This delivers 100% geometric consistency across all 6 degrees of freedom without requiring data augmentation explosion.

// MATHEMATICAL MANIFOLD FORMULATION

Let $\mathcal{M} = \mathrm{SE}(3)$ represent the Lie group of rigid body transformations. The action trajectory $\tau \in \mathcal{T}$ satisfies:

f(g \cdot x, g \cdot p) = g \cdot f(x, p) \quad \forall g \in \mathrm{SE}(3)

Where $x$ is the multimodal sensory observation, $p$ is robot proprioception, and $f$ is the continuous score-matching diffusion network.

MODULE 02

Synthetic Physics Engine & Sim2Real

Collecting physical real-world teleoperation is slow, costly, and fragile. Velocore trains foundation policies in a massively parallel, ray-traced neural physics engine generating 4.2 million interaction steps per second.

  • • 140,000+ randomized physical latents (friction, damping, restitution)
  • • Ray-traced multimodal depth, RGB, and tactile sensor synthesis
  • • Zero-shot transfer to physical hardware without calibration drift
4.2M/s
Physics Steps / Second

Continuous parallel collision detection and contact dynamics.

99.84%
Sim2Real Policy Transfer

Verified across 12,000 novel physical test objects.

±0.04mm
Kinematic Precision

Closed-loop visual and tactile pose reconciliation.

100x
Cost Reduction

Compared to physical teleoperation capture fleets.

MODULE 03

Bare-Metal Tensor Kernels (3.8ms)

Real-world robot dynamics cannot wait for standard 400ms deep learning runtime loops. A falling object or dynamic perturbation requires immediate counter-torque within 5 milliseconds.

Our runtime compiles down to bare-metal C++ matrix kernels with warp-level register reuse and INT4/FP8 quantization, dispatching directly to edge hardware registers in 3.8 milliseconds.

EDGE INFERENCE STACK PROFILER TOTAL: 3.80 ms
1. Multimodal Token Ingestion (RGB-D + Tactile) 1.80 ms
2. SE(3) Diffusion Policy Action Generation 1.40 ms
3. Deterministic Safety Envelope & Register Write 0.60 ms