Skip to main content Table of contents
- Simulation and Environment
- Adam
- Zhihan
- Dimitry
- Version Integration Discussion
- Will
- Referee Gesture
- Zisen
- Environment Issues
- Yuhao
- Framework Discussion
Simulation and Environment
Adam
- Exploring Mujoco
- Josiah suggest Webot simulation as a way to consider
Zhihan
- Successfully transferred V2 to NAO robot
- Ensuring coordination parameters match between simrobot and abstract
- Not yet integrated with Abhinav’s version; plans to merge changes later
- Adding tutorial documentation
- Plan to implement new features:
- Kick skill implementation in V2
- Medium-level skills (e.g., walk to point)
Dimitry
- Facing issue with GUI, yuhao suggest wsl2 version
Version Integration Discussion
- Multi-agent training switch to V2
- Zhihan’s version is focus on transferring and training
- Abhinav’s version focusing on reconfiguration
- Plan to move changes to Abhinav’s branch
- Current issues:
- V2 exhibits unusual behavior with angled kicks
- Strategy: Transfer to Zhihan’s branch first (similar to V1)
- Goal: Achieve stable status combining both versions
Will
- Working on multi-agent reinforcement learning
- Plans to transfer work to V2 in future phases
Referee Gesture
Zisen
- In paper sent by Peter, MobileNet-V2 proven viable on NAO hardware
- Still need for improved efficiency due to concurrent vision pipeline processing
- Insight from paper, number of filters impacts inference time more than layer count
- Tried vanilla GAN for new domain data augmentation
- Josiah suggest also try diffusion for data augmentation
Environment Issues
Yuhao
- Fixed environment bugs, but shared memory has its natural issues
- GNU-related challenges
- Multiple robot simulation causing problems in SimRobot
- BCO implementation pending after merge training
- Josiah suggests focus on single-agent behavior cloning before scaling to multi-agent
Framework Discussion
- RLite: Complex implementation, documentation has latency from code
- Tianshou:
- Undergoing refactoring
- Lightweight design
- Better performance
- CleanRL: Preferred for research purposes