SF Robotics
Discuss a project

Robot learning

Adapt robot behavior to new parts, positions and tasks using demonstrations.

Adapting to new parts and positions

We start from reusable robot skills where they fit and adapt them with demonstrations from your line. For example, a camera-guided station can learn to pick the same part when it arrives rotated or offset in a tote.

How we test an adapted pick

  1. Record successful picks and common misses using examples from your line.
  2. Test on a separate batch with rotated or shifted parts, different spacing and packaging changes.
  3. Record first-pass picks, placement errors, cycle time and operator assists. Review misses before expanding the task.
Record example picks and test new part positions. Measure first-pass picks, placement errors, cycle time and operator assists.

Published research

How Many Demonstrations Does a VLA Need?

Rahul Chand and Miguel Villa Floran compared how six vision-language-action models adapt as the number of demonstrations changes across four simulated task suites.

Read the preprint

SONIC by NVIDIA Research

SONIC shows how a humanoid robot can combine whole-body motion tracking with higher-level plans to walk up to objects and manipulate them.

Sequences from NVIDIA’s SONIC release showing a humanoid robot approaching and handling objects
View the SONIC model diagram
SONIC model diagram showing how motion inputs are encoded and used to control the humanoid robot

Figures 5 and 7 by Luo et al., resized for this page. NVIDIA release · Read the SONIC paper · CC BY-SA 4.0

Discuss an automation project

Contact SF Robotics