Nikos Gkanatsios
Learning-Based Autonomy for Physical and Virtual Agents
Next-Generation Embodied Intelligence • Generative Decision Models • Multimodal Perception
PhD in Robotics, Carnegie Mellon University • Advised by Katerina Fragkiadaki
Autonomy Research & Production Systems (Tesla • NVIDIA)
CV | Scholar | GitHub | LinkedIn | Email
I design learning-based autonomy systems for real-world and virtual intelligent agents through generative decision modeling, multimodal perception, and reinforcement learning.

Research & Publications

My research progresses from representation learning to generative decision intelligence and scalable autonomy learning.

Research Impact Ladder

Perception and Multimodal Representation
Vision-language models for grounded scene understanding.
BUTD-DETR — Vision-language grounding on images and point clouds
ECCV 2022
Project / PDF
ODIN — Unified 2D-3D segmentation model
CVPR 2024
Project / PDF
Spatial Representation and Geometry-Aware Learning
Learning structured spatial representations for reasoning and control.
Analogy-Forming Transformers — 3D in-context learning through relative attention
ICLR 2023
Project / PDF
Act3D — 3D feature fields for equivariant policy learning
CoRL 2023
Project / PDF
Generative Decision Intelligence
Transforming spatial-semantic representations into generative policy and planning models.
ChainedDiffuser — Diffusion planner for long-horizon manipulation tasks
CoRL 2023
Project / PDF
EBM Planner — Energy-based goal generation for long-horizon planning
RSS 2023
Project / PDF
Enabling Scalable Real-World Autonomy
Learning paradigms for next-generation embodied intelligence.
3D Diffuser Actor — Diffusion policy atop semantic geometry-aware representations
CoRL 2024
Project / PDF
Diffusion-ES — Guided diffusion planning for autonomous driving
CVPR 2024
Project / PDF
3D FlowMatch Actor — Scaling 3D policy learning in both capacity and compute efficiency
Preprint
Project / PDF

Older Works

Graph-Structured Semantic Scene Understanding
Grounding Consistency Distillation
ICCV 2021
PDF / Code
Zero-shot Visual Relationship Detection
BMVC 2020
PDF / Code
Attention-Translation-Relation Network for Scene Graphs
ICCV 2019
PDF / Code
The VRD Demo
ICIP 2019
Play live!

Industry Research & Systems Impact

Research and engineering contributions spanning scalable autonomy learning, generative decision modeling, and production deployment of learning-based autonomous systems.

Tesla — Senior Autopilot Machine Learning Engineer Sep 2025 — Feb 2026
  • Contributed to post-training policy optimization strategies using large-scale fleet data for real-world driving adaptation and behavioral refinement.
  • Designed data selection and curation strategies supporting scalable policy learning pipelines.
  • Worked within production-scale training, evaluation, and deployment workflows for autonomous driving systems.
NVIDIA — Research Intern, Robotics & Generative Modeling Jun 2024 — May 2025
  • Developed flow-based generative methods for 3D manipulation policy learning using spatial scene representations.
  • Improved efficiency of 3D policy learning through distributed training optimization and dataset pipeline engineering.
Deeplab — Machine Learning Engineer Jul 2018 — Jul 2020
  • Developed multimodal scene understanding models for graph-structured semantic perception.
METIS Cybertechnology — Artificial Intelligence Engineer Mar 2017 — Jul 2018
  • Developed production NLP and intelligent assistant systems with end-to-end ML pipeline ownership.

Autonomy Systems Engineering Stack

Supporting scalable autonomy learning from research prototypes to production systems.