A few weeks ago, we all met up in person and had a fun, productive, and intense week! Working in three teams, we made significant progress across three fronts: integrating the MuJoCo simulator, simplifying our configurations, and advancing our theory!
All three teams made significant progress during that week and presented their advancements in this meeting. In the end, there is also a fun prize!
0:00 Introduction
1:07 Team 1: Jumanji (Habitat → MuJoCo)
1:34 Demoing Monty Running in MuJoCo
4:39 MuJoCo Integration: Week’s Progress
11:54 Team 2: Something Something Config (Improving Monty’s Configs)
12:57 Configs: Before
13:58 Configs: After
14:59 Config Walkthrough with Tristan
18:52 Config Improvements: Week’s Progress
24:18 Mystery Demonstration by Ramy
25:12 An Interactive TUI to Build and Explore Configs
43:30 Team 3: Three Brains, Two Whiteboards, and a Christmas Tree (Brainstorming)
44:25 What We Set Out To Do & Progress Summary
46:29 A Sample of our New Ideas
48:43 An Alternative to Voting on Locations
53:45 Communicating Displacements Instead of Locations
57:18 Robustness to Deformations by Using Lower-Dimensional Models (T-Shirt/Cable Problem)
1:05:31 Affordances are Causal Links to Behavior Models in Motor Cortex
1:11:49 Distinction Between Goal State and Target Pose
1:13:12 Cortex ↔ Hippocampus
For those who are curious, you can find the MuJoCo integration prototype here: wip: MuJoCo Simulator Prototype by jeremyshoemaker · Pull Request #853 · thousandbrainsproject/tbp.monty · GitHub
The Hydra config simplification is already merged into tbp.monty and was summarized by @tslominski in this thread: We changed how configurations work (again)
In the next weeks, we will release research meeting recordings that go into the new theoretical ideas in more depth 