The Road to Autonomy - Unlocking Humanoid Loco-Manipulation
Autonomous humanoid robots need to walk, balance, and manipulate objects - often all at once.
But the approaches that work for locomotion look surprisingly different from those that work for manipulation, and combining them into coherent whole-body behaviour remains one of robotics' hardest open problems.
This talk traces the path toward humanoid loco-manipulation, covering from vision-based locomotion to recent whole body control methods like TWIST and SONIC that attempt to bridge the gap.
We'll look at how learning-based approaches differ from classical control pipelines like those behind Boston Dynamics' robots, not just in results but in how quickly they generalise across new platforms and tasks.
The talk further touches upon learning for locomotion vs manipulation and ends with comparing one step techniques to multi-step diffusion/flow matching and some interesting open questions.
About the speaker:
Tanmay Agarwal is a research engineer at Skild AI, working on whole-body control and policy learning for humanoid robots.
Skild is building general-purpose foundation models for robotics. Earlier, he worked at Carnegie Mellon's Robotics Institute on mobile manipulation and perception for healthcare robotics.
He previously built robotics systems at Addverb Technologies in India and led robotics at the ZINE robotics lab at the National Institute of Technology, Jaipur.
X: tanmayagarwal98
To attend online:
Add to calendar: https://shorturl.at/30Bkr
Gmeet link: meet.google.com/tun-fyip-srr
Look forward to seeing you!