About the role
Responsibilities:
- Adapt open-weight foundation models (OpenVLA, Octo, 0) for multi-modal quadruped navigation and behavioral manipulation.
- Formulate multi-modal tokenization schemes that translate visual observations and natural language instructions into high-level spatial waypoints and behavior tokens.
- Direct parameter-efficient fine-tuning (PEFT/LoRA/OFT) pipelines on downstream surveillance datasets.
- Architect the multi-task vision trunk incorporating proprietary security analytics (threat detection, facial recognition) into the VLA representation layer.
Requirements:
- Ph. D. or master's in computer science, artificial intelligence, or robotics.
- Minimum 4 years of experience training Large Multimodal Models (LMMs), Vision Transformers (ViT), or robotic imitation learning policies.
- Technical Strengths: PyTorch, JAX, Hugging Face, OpenVLA, Octo, Vision Transformers (ViT), Cross-embodiment data (Open X-Embodiment).