Robotics

The challenge

Industrial environments — manufacturing floors, hazardous facilities — demand robots that can perceive, reason, and act with full data sovereignty:

• ▲ All sensor feeds, operational logs, and inference outputs stay on-site
• ▲ No cloud upload, no third-party exposure, no IP leakage
• ▲ Equipment is non-standardised, lighting varies, and a wrong action can be irreversible

Classical manipulation pipelines break under distribution shift. Large foundation models require cloud-scale inference. Compact models generalise but offer no calibrated uncertainty signal for safe operation — they act with equal confidence in familiar and unfamiliar scenes alike.

PRAESC closes this gap with systems that combine edge-deployable perception, principled uncertainty quantification, and a confidence gate that keeps humans in the loop precisely when it matters.

Our approach

• ▲
  Compact VLA perception. Sub-1B parameter Vision-Language-Action models running on a Nvidia Jetson-class device. Semantic scene understanding, reasoning, and planning without cloud inference.
• ▲
  Physics-grounded Bayesian confidence. A GPU-efficient Bayesian kernel that compares sensory input signals against the robot's own dynamics model, updating its beliefs about the world. This produces a calibrated confidence score about its actions; the system knows what it doesn't know.
• ▲
  [Coming-up] Confidence-gated execution. Above threshold: the robot acts. Below: it pauses, generates a plain-language explanation, and holds for operator confirmation. Human oversight exactly where uncertainty is high.
• ▲
  Simulation-first development. Full digital twin in Sim before any real hardware. Sim-to-real gap closed iteratively via a structured feedback loop from on-site failure cases.