Responsible for conducting SoTA research across multiple directions, including
RL-based planning
- Worked closely with ML infra team to define and implement RL training infrastructure
- Developed core off-policy RL training algorithm used by RL planner
- Researched multiple approaches to integrate safety constraints into RL planning system
Generative approach to behavior prediction and Imitation Learning
- Did comprehensive analysis of generative modeling techniques as an alternative to regression-based Imitation Learning.
- Developed and implemented diffusion-based generative behavior prediction model.
- Generalized diffusion-based prediction model to provide interaction-aware joint prediction, showed superior performance on internal dataset for joint prediction.
- Developed novel diffusion distillation approach to reduce inference latency of diffusion-based predictor.
Anomaly detection for behavior models
- Implemented EBM-based anomaly detection model and showed superior performance compared to heuristic-based method used previously.