FOUNDING PROJECT / IRVL @ UT DALLAS
VLA-Replica
A replicable real-world benchmark for vision-language-action models using the low-cost SO-101 arm.
OVERVIEW
The project that started the network.
VLA-Replica was created to make real-world robot manipulation evaluation easier to reproduce and compare. It uses an affordable, off-the-shelf SO-101 arm and a carefully specified physical setup.
RobotReplica extends that idea beyond one benchmark: partner organizations maintain evaluation sites for different robot embodiments so policies can be tested on matched, real hardware.
BENCHMARK DESIGN
Consistent tasks.
Meaningful variation.
The benchmark standardizes the robot, workspace, task objects, camera views, demonstrations, and evaluation scenes while testing both familiar and shifted conditions.
EVALUATION
Evaluate on the UT Dallas SO-101 site.
Researchers with SO-101-compatible manipulation policies can contact RobotReplica to discuss interfaces, checkpoints, task coverage, and evaluation reporting.