🏢 Industry Internship Projects

🤖 BiDexVLA: Hybrid Bimanual Dexterous Grasping Framework

Company: Samsung Research China
Duration: 2025.05 - 2025.10
Project Homepage: https://sevenfo.github.io/BiDexVLA/

Background & Objectives

Developed a hybrid framework combining Vision-Language-Action (VLA) models with traditional planning for bimanual robot grasping.

Key Contributions

  • Designed a two-stage architecture: model-based pre-grasp generation followed by a predict-refine diffusion model fusing vision, force feedback, and point cloud data.
  • Built a bimanual teleoperation data collection system with automated trajectory generation.
  • Fine-tuned multimodal LLMs (e.g., Qwen-2.5-VL) for semantic planning.

Key Achievements

  • 41.8% increase in grasping success rate and 23.5% reduction in execution time compared to baselines.
  • Paper “BiDexVLA: A Hybrid Framework for Fast and Robust Bimanual Dexterous Grasping” submitted to ICRA 2026 (Under Review).

Technical Highlights

  • Multimodal perception fusion.
  • Hierarchical planning.
  • Real-time bimanual coordination.

📝 CARE: Context-Aware Retrieval-Enhanced Reasoning

Company: MetaGPT
Duration: 2024.06 - 2024.12
Project Homepage: https://foundationagents.github.io/CARE/

Background & Objectives

Contributed to research on improving LLM context fidelity through a native retrieval-augmented reasoning framework.

Key Contributions

  • Designed the CARE framework to handle context hallucination with native retrieval.
  • Implemented two-phase training: supervised fine-tuning and curriculum-based reinforcement learning.
  • Developed a reward function using Group Relative Policy Optimization (GRPO).

Key Achievements

  • +15.29% average F1 improvement over baselines, with higher gains on complex datasets.
  • +13.69% improvement on counterfactual benchmarks.
  • Paper “Improving Context Fidelity via Native Retrieval-Augmented Reasoning” accepted at EMNLP 2025 (4th author).

Technical Highlights

  • Curriculum learning for multi-hop reasoning.
  • Evidence retrieval across model scales.
  • Performance in long-context QA.

🛠️ Aviation Big Data Natural Language Modeling Tool POC

Company: MetaGPT
Duration: 2024.06 - 2024.12

Background & Objectives

Participated in POC validation for a natural language tool for aviation data analysis using multi-agent systems.

Key Contributions

  • Validated 7+ scenarios such as anomaly detection and path planning.
  • Produced 15+ analysis documents.
  • Built baselines with statistical, deep learning, and operations research models.
  • Developed multi-agent prompt engineering.

Key Achievements

  • Demonstrated natural language interface feasibility for aviation analytics.
  • Implemented models for pattern recognition and optimization.
  • Created evaluation framework.

Technical Highlights

  • Multi-agent collaboration.
  • Prompt engineering for aviation tasks.
  • Integration of modeling approaches.

🔬 Research Projects

Large Model-based Robot Visual Manipulation and Collaboration Methods Research

Institution: Beihang University
Duration: 2024.01 - Present
Funding: Siyuan Alliance Fund

Background & Objectives

Developed a multimodal multi-agent framework and imitation learning models for robot manipulation and collaboration.

Key Contributions

  • Designed multi-agent task planning with visual observation, reflection, decision-making, skill management, and feedback.
  • Built simulation environment in Isaac Sim for lunar base scenarios.
  • Collected demonstration data and trained skill models using ACT, Diffusion Policy, RT-1, and OpenVLA.

Key Achievements

  • Improved manipulation capabilities in simulated environments.
  • Optimized imitation learning for skills.

Technical Highlights

  • Multimodal fusion.
  • Hierarchical planning.
  • ROS integration.

Embodied Intelligence Aircraft Online Autonomous Motion Planning Research

Institution: Beihang University
Duration: 2023.09 - 2024.09
Funding: CALT University-Industry Collaboration Fund

Background & Objectives

Implemented LLM-based planning for UAV autonomous motion in open environments.

Key Contributions

  • Built point cloud semantic segmentation using OWLv2, SAM, and XMem.
  • Applied prompt engineering with DeepSeek-coder for code generation.
  • Integrated Voxposer architecture with CoppeliaSim/AirSim/Isaac Sim and ROS.
  • Added feedback for dynamic planning.

Key Achievements

  • Enhanced adaptability and task success rates.
  • Open-sourced components.

Technical Highlights

🎓 Academic & Early Projects

Distance Metric-based Meta-Learning Few-Shot Classification Method Research

Duration: 2023.01 - 2023.06
Type: Undergraduate Thesis

Background & Objectives

Proposed Graph Prototype Network (GPN) for few-shot classification using graph attention mechanisms.

Key Contributions

  • Designed GPN architecture.
  • Implemented PyTorch pipeline.
  • Conducted experiments and visualization.

Key Achievements

  • Over 20% accuracy improvement vs. baselines.
  • 69.4% on miniImageNet 5-way 5-shot.

Technical Highlights

  • Metric and meta-learning.
  • Graph attention networks.

Multimodal Deep Learning-based Bus Driving Safety Evaluation Research

Duration: 2021.07 - 2022.06
Type: Municipal-level University Student Innovation Project

Background & Objectives

Developed system for detecting safety-related sound events in buses.

Key Contributions

  • Designed CRNN model.
  • Deployed with ONNX C++.
  • Built Qt UI.

Key Achievements

  • Rated “Good” on completion.
  • Functional detection system.

Technical Highlights

Hexapod Biomimetic Robot

Duration: 2020.07 - 2021.06
Type: University-level SRTP Project (Project Leader)

Background & Objectives

Led development of hexapod robot with kinematics and navigation.

Key Contributions

  • Modeled kinematics.
  • Implemented ROS navigation.
  • Developed Flask web interface.

Key Achievements

  • Rated “Good” on completion.

Technical Highlights

  • Kinematic modeling.
  • ROS and web development.

🏆 Competition Experience

National Smart Car Competition

  • 2020: National First Prize 🥇
  • 2021: National Second Prize 🥈

Competition Content: Path planning, perception, and control for intelligent vehicles.

For more project details and code implementations, please visit my GitHub 🚀