Check out the paper and feel free to reach out! Incredibly grateful to my collaborators Lokesh and Nikhil at the Dynamics Robotics and Control Laboratory (DRCL), USC. Special thanks to Prof. Quan Nguyen for his valuable inputs throughout the project! (6/n)
Policies may get stuck in local optima by choosing an undesirable sequence of modes. To mitigate this, we introduce a task-agnostic mode-switching preference reward using mode ranks specified by the user. (5/n)
We propose an oracle-guided policy optimization framework leveraging a hybrid automata perspective to design multi-mode oracles. This abstraction results in structured exploration using a single oracle across different tasks and robots. (4/n)
Dynamic loco-manipulation calls for whole-body and contact-rich control: a hard exploration problem for deep RL due to susceptibility to local optima. Current approaches rely on:
- Task-specific reward shaping
- Multiple low-level/skill policies (3/n)
Introducing an oracle-guided policy optimization framework for synthesizing RL policies to tackle dynamic loco-manipulation
✅ Single multi-mode policy per task
✅ One oracle, same reward weights & hyperparameters across different robots & tasks
(2/n)
Excited to share our recent work:
Dynamic Bipedal Loco-manipulation using Oracle Guided Multi-mode Policies with Mode-transition Preference
Website: indweller.github.io/ogmplm/
Preprint: arxiv.org/abs/2410.01030
Video: youtu.be/gfDaRqobheg?...
🧵(1/n)
