- Domain
- Robotics / learned behavior
- Period
- 2026—present
Context
Learned policies do not arrive at hardware ready to run. Demonstrations, observations, controller timing, calibration, and safety all have to agree; a failure at any seam changes what the policy can actually do.
Contribution
Built the path from Apple Vision Pro teleoperation and synchronized episode capture through VLA policy adaptation to interpolated, safety-constrained execution on a Franka Research 3. The work spans data collection, training inputs, and deployment rather than one isolated model step.
Evidence
The complete path runs on physical hardware. The controller’s 1 kHz requirement is treated as an operating constraint—not a result—and bridged from slower policy inference with smoothing and runtime checks.