Jeffery

  • Discussed dataset approach with Zisen

  • BHuman dataset: select 2-3 most important images per second for training

Zisen

  • Confirmed BHuman logic working in wistex system

  • Referee thread: BHuman loads move net and heuristic logic

  • Questions raised:

    • Entry point for running onnx?

      • Yuhao’s suggestion: Trace BHuman code for onnx usage
    • Number of trajectories needed for IL algorithm?

      • Yuhao’s response: As many as possible
  • Note: Referee thread needs to infer onnx after whistle

Yuhao

  • Performance issues with py-log-reader

    • Parsed first data batch (1.39GB) in 8 hours
  • BHuman2024 merge failed due to skill implementation changes:

    • Skills changed from separate classes to CABSL options under SkillBehaviorControl

    • Uses reflection mechanism for data storage

  • Zhihan’s question: Location of reflection data storage?

    • Yuhao’s response: Possibly in SkillBehaviorControl or blackboard (uncertain)
  • Will share skill change details in Slack

Edisen Shen

  • Investigating Walkengine

  • Features walk phases with interpolated step position and velocity

  • Adam’s question: Control of desired angle at each time step?

    • Response: Yes, takes speed request, sends desired angle

Adam

  • Received assistance from mechanical engineering team

  • Progress with mujoco:

    • Nearly achieved good standing behavior

    • Planning to use Deepmind soccer env for training

Zhihan

  • Discussed log usage with Yuhao

  • Haven’t implemented py-log-reader infrastructure

  • Working with Siddhant on UT-Austin undergrad recruitment

  • Focus on data collection and 2024 codebase transfer

  • William’s question: Status of abstractSim policy deployment

    • Response: Main challenge is action gap between simRobot and real robot

    • Working on neural network solution to bridge walk engine dynamics

Siddhant

  • Working on policy patching

  • Removing game controller

  • Peter’s question: Whether to remove GameController or modify for constant playing state

    • Response: Will evaluate easiest modification approach

Abhinav

  • Working on abstractSim v2

    • Improved user understanding

    • Collaborating with Zhihan on SimRobot to abstractSim V2 migration

Parv

  • Examining goalkeeper code simulation

  • Studying RL implementation for goalkeeper scenarios

Zhihan

  • BHuman Fast showing unusual behavior

    • Robot reported as broken

    • Adam: Confirmed similar previous observations