GUIDE is built upon our recent platform CREW. Please check it out as well. Many has started using CREW in their research!

generalroboticslab.com/CREW/

Website with video, code, and paper: generalroboticslab.com/GUIDE
Duke' press release: pratt.duke.edu/news/trainin...

Read the Full Paper Dive into GUIDE’s technical details and our exciting findings. Explore the future of human-in-the-loop learning! Kudos to our team Lingyu Zhang, Zhengran Ji, and Nicholas Waytowich!

Human Cognitive Tests: We didn’t just study AI - we looked at humans too! Cognitive tests revealed how individual differences impact training outcomes, helping us understand how diverse skills translate to better AI guidance. #Neuroscience #HumanFactors

How Does It Perform? Our experiments show GUIDE delivers 30% higher success rates than baseline RL models in complex tasks like navigation and multi-agent hide-and-seek. And with just 10 minutes of human feedback, it surpasses previous methods by up to 40%.

Mimicking Human Feedback: After initial training, GUIDE keeps improving! While the human is providing feedback, we train a feedback model that replicates human guidance, allowing the agent to continue learning independently. This reduces human effort and ensures robust, ongoing performance gains.

Continuous Human Guidance: GUIDE’s interface allows continuous feedback - no more simple “yes/no” or “good/bad” labels. Trainers can provide nuanced guidance at every decision step, making learning more natural and expressive for both AI and trainers.

Largest Human Studies Ever: Most studies on human-guided AI have involved fewer than 10 participants - often the authors themselves! GUIDE stands apart with the largest human subject study to date in this field, involving 50 participants to rigorously validate our approach.

Why GUIDE? Real-time decision-making is a tough nut to crack for AI, especially in high-stakes tasks. GUIDE leverages human feedback in real-time, grounding it into _dense rewards_ to accelerate AI learning, even in challenging environments with sparse feedback signals.

That’s amazing! Congratulations!!!

Just joined the platform. Would love to join the list if possible!