2021 Blue Sky Session 2

In this talk

1. 0:19

   Causal reasoning for safer navigation

   Kayleen Stocking contrasts a pedestrian waiting at a bus stop with someone near a soccer ball who might run to retrieve it. Causal graphical models could connect objects and possible intentions, helping a robot judge when passing is safe even in situations absent from its training data. She says model generation and dynamic reasoning remain open challenges, and suggests causal models could complement reinforcement learning.

2. 7:44

   A shared cloud for robot research

   Victoria Dean of Carnegie Mellon proposes a remotely accessible center with real robots, shared tasks, benchmarks, and public experiment data. The goal is to reduce setup work and make real-world robotics research accessible to people without hardware expertise. She outlines challenges including robot maintenance, task selection, safe access, funding, and scaling; the proposed plan aimed for 100 robots and 400,000 robot hours over two years.

3. 13:30

   Separate data collection from learning

   Martin Riedmiller of DeepMind presents Collect and Infer, a framework that treats gathering useful experience and extracting policies from stored data as linked but distinct problems. He argues that both the collection strategy and the inference algorithm can strongly affect policy quality under a fixed data budget. Scheduled Auxiliary Control is one approach, using a scheduler to choose skills that generate useful data.

4. 19:18

   Robots supporting robot learning

   Flo describes Dual-Arm Adversarial Robot Learning, or DARN, as a way to reduce the human work involved in resets, reward computation, and safety during manipulation experiments. A trainer robot can reposition objects or introduce challenges while a trainee learns; known object locations can also support safety checks and perception datasets. The added robot brings cost and complexity, and the discussion compares this lab-scale approach with Dean’s shared robotics cloud.