Effective Use of Human Data
Happy Thanksgiving! Would you have liked an AI preparing your feast? Or maybe cleaning up afterwards? On the road to creating collaborative AI agents that can better understand and assist humans, Optimal Behavior Prior presents an efficient way to use data collected from humans using the Overcooked-AI benchmark environment.
Figure Rework
Through FAR.AI, I had the opportunity to design communications for the paper. The main figure was a centerpiece highlighting the key concepts from the paper and an algorithm also clarified the process.
While the original figure was already well-done, we explored other ways to strengthen the ideas by:
- Stacking the environments to show a given environment with agents in greater detail.
- Combining the policy, agent, and neural network model images together into a single image emphasized the color & labels more clearly.
- Simplifying the human-data image.
- Adding a narrative storyline with bold keywords to highlight the discrete steps.
- Incorporating subtext from the algorithm underneath keywords added to the explanation.
Poster Design
After some additional back & forth to finalize text description and LaTex labels, the newly redesigned figure and other key elements from the paper were extracted to create the resulting poster. Normally, I use Adobe Illustrator or InDesign for layout projects. However, in this case, figures were formatted in Figma then printed on a 2’x3’ poster.
Check out the full paper, “Optimal Behavior Prior: Data-Efficient Human Models for Improved Human-AI Collaboration” presented at the NeurIPS 2022 Human in the Loop Learning (HiLL) Workshop. Special thanks to Micah Carroll for providing the initial figures, sample poster templates, and constructive feedback throughout the entire process.