Our representative, Yuta Oriike, presented a poster at the iHR workshop

On September 19, 2026, our representative Yuta Oriike gave a poster presentation at the 15th meeting (iHR15) of the Intelligent Home Robotics (iHR) technical committee of the Robotics Society of Japan, held at Soka University on September 19–20.

The poster, titled “Discovering reward hacking in RL Token replication experiments, and a reward design loop with humans and an AI agent (VLM judge),” was presented as an interim report.

We independently reimplemented RL Token, a robot learning method, on a simulated task of moving three oranges onto a plate. Human review of the videos revealed that a metric which seemed to improve was measuring how often an orange came near the plate, not how often it was placed on it.

The presentation described how this finding led us to rebuild the success criteria and rewards, and how work was divided between people, an implementing AI agent, and an AI that checks images. The image-checking AI (VLM judge) agreed with human judgments 98.6% of the time, suggesting that combining people and AI agents can make revising evaluation and rewards more efficient.

At the time of the presentation, we had verified that the revised reward is calculated correctly. Improving performance through learning remains future work.

The presentation was also introduced in a note article by the Nakazawa Laboratory at Kanazawa Institute of Technology, where Oriike is a visiting researcher. The research article below describes the work in detail.