ICRA Keynote 4 - Manipulation, Humanoids, Embodied Design

In this talk

1. 0:47

   Dexterous hands and object-centered learning

   Janette Buck argues that multi-fingered hands can enable faster tool use and finer control than two-finger grippers, though they are difficult to operate. Her team trains a single reinforcement learning policy in simulation to move objects toward poses drawn from human video, without copying human hand motions. Fine-tuning that policy for insertion tasks improves precision and recovery, but relies on rigid-object pose tracking that can fail.

2. 21:41

   Building humanoids with humanlike bodies

   Kento Kawaharazuka describes tendon-driven robots designed with flexible spines, redundant motors, soft surfaces, and expressive features. Examples include Kengoro, with 124 motors and evaporative cooling, and modular Musashi robots that can perform tasks such as driving and table setting. He sees humanlike bodies as one path toward safer physical interaction and better communication, while acknowledging that simulation and control remain difficult.

3. 43:28

   Modular robots for changing missions

   Nikos Sagaris explains how plug-and-play modules can be assembled into robots for construction, space infrastructure, and other demanding settings. A construction platform switched configurations in about 10 to 12 minutes, and a faulty joint was replaced in three or four minutes. He also describes designs that reduce energy use and research into robots that adapt their movements when actuators degrade.

4. 1:02:35

   Foundation models trained on varied data

   Professor UK of UT Austin and NVIDIA presents humanoid models that combine language and vision inputs with action generation, alongside whole-body controllers. His team uses motion capture and egocentric videos to build robot skills, reporting improved performance as video pretraining grew from 1,000 to 20,000 hours. He also describes DreamZero, which predicts future video and robot actions, and argues that evaluation should include reliability, uptime, and throughput alongside task success.

5. 1:20:28

   Open questions on generalization and reliability

   In the discussion, Professor UK says post-training and reinforcement learning may help raise success rates, but deployment needs broader measures than success alone. Diverse pretraining videos can help models generalize even when they differ from the eventual job. He describes combining modalities and learning from failures as unresolved research challenges.