Cup Stacking Robot - Embodied AI Hackathon

Overview

Developed an autonomous cup stacking system using NVIDIA’s GR00T 1.5 foundational model and Action Chunking Transformer (ACT) for the SO-101 robotic arm. This project was completed during the Embodied AI Hackathon hosted by Seeed Studio and NVIDIA in October 2025.

Cup Stacking Demo
Autonomous Cup Stacking in Action

Project Highlights

Fine-Tuned Foundational Models

  • NVIDIA GR00T 1.5: Leveraged NVIDIA’s cutting-edge foundational model for robotic manipulation
  • ACT (Action Chunking Transformer): Implemented imitation learning policy for precise grasp-transfer-stack sequences
  • Model Deployment: Optimized for real-time inference on NVIDIA Jetson Thor edge AI platform

Data Collection & Training

  • Teleoperation Demonstrations: Collected 80+ high-quality demonstrations of cup stacking sequences
  • Imitation Learning: Trained policy to replicate human demonstrations with robust generalization
  • Sequence Learning: Mastered multi-step manipulation: detect → grasp → transfer → stack

Edge AI Deployment

  • Platform: NVIDIA Jetson Thor for on-device inference
  • Performance: Achieved robust real-time control with minimal latency
  • Integration: Seamless deployment from training to production on edge hardware

Technical Implementation

The system architecture combines:

  • Robot Platform: SO-101 robotic arm with 6-DOF manipulation
  • Vision System: Orbbec camera for object detection and pose estimation
  • Control Framework: ROS2-based control with micro-ROS integration
  • Motor Control: Feetech servos with precise position control

Results

Successfully demonstrated autonomous cup stacking with:

  • Consistent grasp detection and execution
  • Smooth transfer motions minimizing cup disturbance
  • Reliable stacking with proper alignment
  • Real-time performance on edge hardware

Challenges & Learnings

  • Data Quality: Importance of diverse demonstrations for robust policy learning
  • Edge Optimization: Balancing model complexity with inference speed on Jetson Thor
  • Hardware Integration: Coordinating vision, control, and manipulation in real-time

Team & Acknowledgments

Team Members: Narendhiran Saravanane, Anirudh Manjesh, Ayan Syed, Josue Tristan

Hackathon: Embodied AI Hackathon Organizers: Seeed Studio & NVIDIA Date: October 2025

Links & Resources

This project demonstrates the power of combining foundational models with imitation learning for robotic manipulation tasks, deployed on edge AI hardware for real-world applications.

Cupstacking | @narendhiran2000