AppliedIntuitionResearch/nord
AppliedIntuitionResearch/nord is a 3 billion parameter Vision-Language-Action (VLA) model developed by Applied Intuition Research, designed for autonomous driving. It directly predicts driving trajectories as discrete tokens without requiring chain-of-thought reasoning, achieving high performance on the NAVSIM PDMS benchmark (0.8626). This model is notable for its data-efficient training using Supervised Fine-Tuning (SFT) and Dr. GRPO, utilizing 3x fewer tokens than reasoning-based VLA baselines, making it suitable for real-time driving applications.
Loading preview...
Overview
AppliedIntuitionResearch/nord is a 3 billion parameter Vision-Language-Action (VLA) model developed by Applied Intuition Research, specifically engineered for autonomous driving tasks. This model distinguishes itself by directly predicting driving trajectories as discrete tokens, bypassing the need for explicit chain-of-thought reasoning. It leverages a data-efficient training approach combining Supervised Fine-Tuning (SFT) with Dr. GRPO, enabling it to achieve high performance with significantly less training data.
Key Capabilities
- Direct Trajectory Prediction: Predicts
(x, y, heading)trajectories at 10 Hz over 4 seconds directly from visual and contextual inputs. - Data Efficiency: Achieves strong performance using 3x fewer tokens compared to reasoning-based VLA baselines, making it efficient for training and deployment.
- High Driving Performance: Demonstrates a NAVSIM PDMS score of 0.8626, indicating robust performance in simulated driving environments, with high scores in no-at-fault collisions (0.9737) and drivable area compliance (0.9522).
- Vision-Language-Action Integration: Processes camera inputs, ego-velocity, and driving commands to generate precise trajectories.
Good For
- Autonomous Driving Research: Ideal for researchers and developers focused on efficient and direct trajectory prediction in autonomous systems.
- Real-time Driving Applications: Its non-reasoning approach and data efficiency make it suitable for scenarios requiring fast and reliable trajectory generation.
- Benchmarking VLA Models: Provides a strong baseline for evaluating data-efficient VLA architectures in driving simulations.