MIND MELD: Personalized Meta-Learning for Robot-Centric Imitation Learning
summary
MIND MELD meta-learns a personalized embedding via mutual information and a Bi-LSTM to map suboptimal human corrective labels to estimated optimal labels, trained using calibration tasks with RRT and MPC ground truth in an AirSim driving simulator. A synthetic test shows 55% improvement in corrective labels, a human study uses 76 calibration participants and 42 test participants, and the system yields near 100% probability* of reaching the goal and statistically significant gains (p < .001).
AI-written summary · about
chapters (6)
- 0:00Introduction and motivation
- 1:38Robot-centric learning challenges
- 3:10MIND MELD architecture overview
- 5:03Human-subjects study design
- 7:14Experimental results and findings
- 9:27Future work and acknowledgments
related talks
28mPreference Learning from Minimal Human Feedback for Interactive
50mFlexible Neural Networks and the Frontiers of Meta-Learning
01h 35mDeep Robotic Learning
22mA Hybrid Meta-Learning Framework for Adaptive Safe Controller Synthesis of Dynamical(…)
03mIdentifying Children Metacognitive Monitoring Performance Through Facial Expressions
40mTeaching a Car to Drive Itself by Imitation and Imagination
13mMEDebiaser: A Human-AI Feedback System for Mitigating Bias in Multi-label Medical Image Classific...
32mAn Introduction to 'metalearners', a Flexible MetaLearner Library in Python