Researchers led by Apoorva Vashisth and Manav Kulshrestha have built RePlan-VLA, a framework that lets vision-language-action models carry out long tasks made of several dependent steps, according to a paper on arXiv. Such models handle short manipulation well but tend to fail as tasks get longer.

Rather than asking the model itself to plan, remember and recover from errors, RePlan-VLA watches the robot's motion to judge progress and revises the rest of the plan according to what actually happened. It works with any such model and needs no extra fine-tuning.

On the RoboCasa365 and CALVIN benchmarks it raised success rates by 21.9% and 10.4% over the next best method, and the team also tested it on a real Franka Research 3 robot arm.

Source: arXiv Robotics (cs.RO)

Background

Apoorva Vashisth

Robotics researcher

Robotics researcher.