Team: We are Team 6: Voltage Drop.

Members:
th31dkk
Guitar-Fann
Voidless_Psychoo

Summary: We trained an AI powered robot using a Hugging Face LeRobot SO-101 to pick up a mini baseball bat and carefully aim and hit a mini baseball off of a stationary tee, into a 3d printed goal, which will be moved every trial. The AI arm will then detect if it scored or missed and strike a celebrating pose for success, and frustation pose for failure.

Our mission was to train a robotic arm to be able to consistently perform a task that requires precision and adaptation to several different factors.

Although demonstrated with a simple model, our mission can be used in a multitude of applications where robots will have to pick up tools and hold them and correct errors, and.

• Our unique approach is that the robot is able to adapt on factors that weren't really refined on during training. When the bat was placed off of the normal placement, the robot succeeded in recovering and using the bat. We also incorporated visual and audio feedback for successful runs.
• Our innovative approach is having markers on the field for location of the ball and the bat so that it is easier to reset runs accurately.

• During Teleop/Recording Mode, we repeated the same action repeatedly to ensure good training data for the AI to use. Teleoperation Demonstration
• For training, we used the ACT model with 30 episodes and 15k steps to reduce loss as much as possible in the short amount of time.
• While the inference evaluation is not perfect, is is ablt to complete the task effectively for a significant amount of the runs.
  • Video of inference eval>gt;

• Our current implementation can be generalized to other tasks that require simple manipulation of a stationary object.
• It can be adapted to complete tasks such as knocking down bowling pins and using tools like hammers.
• Commands necessary to control the robot are the 'lerobot-record' command in the command line with parameters such as the HuggingFace Model and Dataset. The external ESP-32 hardware is programmed using the Arduino IDE.

• Task Completion: https://youtu.be/UVUJsXMg3_4
• HuggingFace Dataset: https://huggingface.co/datasets/Guitar-Fan/ballhit2-record-test
• HuggingFace Model: https://huggingface.co/Guitar-Fan/act_movegoal_10ksteps

This is the directory tree of this repo, you need to fill in the mission directory with your submission details.

AMD_Robotics_Hackathon_2025_ProjectTemplate-main/
├── README.md
└── mission
    ├── code
    │   └── 
    └── wandb
        └── 
ur training job>
b>

$ tree outputs/train/smolvla_so101_2cube_30k_steps/wandb/
outputs/train/smolvla_so101_2cube_30k_steps/wandb/
├── debug-internal.log -> run-20251029_063411-tz1cpo59/logs/debug-internal.log
├── debug.log -> run-20251029_063411-tz1cpo59/logs/debug.log
├── latest-run -> run-20251029_063411-tz1cpo59
└── run-20251029_063411-tz1cpo59
    ├── files
    │   ├── config.yaml
    │   ├── output.log
    │   ├── requirements.txt
    │   ├── wandb-metadata.json
    │   └── wandb-summary.json
    ├── logs
    │   ├── debug-core.log -> /dataset/.cache/wandb/logs/core-debug-20251029_063411.log
    │   ├── debug-internal.log
    │   └── debug.log
    ├── run-tz1cpo59.wandb
    └── tmp
        └── code
g
    ├── run-tz1cpo59.wandb
    └── tmp
        └── code
   └── code