A new paper on arXiv introduces PACER, a method for learning vision-language-action (VLA) policies for robotic disassembly. The authors note that disassembling end-of-life products is challenging because robots must accommodate product variation, follow prerequisite sequences, and approach components precisely for contact-rich manipulation. The abstract argues that these requirements go beyond what standard supervised fine-tuning (SFT) can provide.
PACER instead learns from human corrections, allowing the robot to adjust its behavior during task execution. The "process-aware" aspect means the policy is designed to respect the necessary order of operations in disassembly, rather than treating each action independently. This is a key departure from conventional SFT, which typically relies on static demonstrations.
The abstract is truncated, so details on experimental results or comparisons are not available in the source. However, the framing suggests that PACER targets the gap between generic VLA training and the structured, sequential nature of real-world disassembly tasks. The work appears aimed at making robotic disassembly more adaptable to the variability of end-of-life products.