Skip to content
Vestval

Robotics Lab

Computer Vision

Perception systems that make sense of cluttered, changing real-world scenes well enough to act on them safely.

Overview

Computer vision is the sensory foundation of every capable robot. The lab's work focuses on perception that is robust to lighting, clutter and novelty — because deployed environments never look like the lab.

Vision is developed jointly with action: perception is only useful if it survives contact with the closed loop.

What this covers

Robust detection

Detection and segmentation that hold up under occlusion, distortion and unusual objects.

Depth & geometry

Metric depth and geometry from mixed sensor stacks.

Tracking under uncertainty

Multi-object tracking that reasons about occlusion and re-identification.

On-device inference

Models sized and quantized to run on real robot compute budgets.

Domain adaptation

Transferring lab-trained perception to deployment environments.

Failure awareness

Perception that signals uncertainty rather than confidently hallucinating.

How it works

  1. 1

    Data is collected in target environments with instrumented protocols.

  2. 2

    Models are trained and evaluated on domain-relevant benchmarks.

  3. 3

    Deployment pipelines quantize and monitor on-device performance.

  4. 4

    Failure signals are surfaced and fed back into the training loop.

Use cases

Manufacturing pilots

Bin picking and quality inspection under variable lighting.

Perception that survives factory floor conditions.

Service robotics research

Robots operating in human environments.

Reliable perception across cluttered, changing scenes.

FAQ

Frequently asked

  • Both — hybrid stacks often win on latency and reliability.