The secret behind IIIT Hyderabad’s collision-free autonomous driving

IIIT Hyderabad has developed an electric self-driving car designed to autonomously navigate point-to-point routes with advanced collision avoidance mechanisms. Equipped with state-of-the-art technologies, including 3D LIDAR, depth cameras, GPS, and an Attitude and Heading Reference System (AHRS), this vehicle can interpret and execute open natural language commands to reach specified destinations.

Professor Madhava Krishna, Head of the Robotics Research Centre and the Kohli Centre for Intelligent Systems (KCIS) at IIIT Hyderabad, sheds light on the unique capabilities of IIIT Hyderabad’s self-driving car and the groundbreaking research on driving it.

The car uses SLAM-based point cloud mapping to create a detailed representation of its surroundings, enabling real-time localisation during movement. Its trajectory optimisation framework, inspired by Model Predictive Control, ensures seamless path planning and execution.

This framework integrates data-driven models to optimise decision-making, delivering faster and more accurate navigation. The research behind this innovation has earned recognition at prestigious global conferences, marking a milestone in autonomous vehicle technology.

REDEFINING NAVIGATION WITH OPEN SET COMMANDS :

Human navigation often relies on contextual clues and verbal instructions, such as "turn right at the white building" or "stop near the park entrance." IIITH’s car mirrors this human-like approach by incorporating open-set navigation, which enables it to navigate using real-world landmarks.

Unlike conventional methods that rely on high-resolution GPS or detailed High-Definition (HD) maps, which are resource-intensive, IIITH’s approach utilises lightweight open-source topological maps, like OpenStreetMaps (OSM).

While these maps are metrically less accurate, IIITH enhances their usability by integrating them with contextual language landmarks, such as “under the bridge” or “near the playground.” This technique enables the car to navigate dynamically to new locations without explicit prior training, offering remarkable zero-shot generalisation capabilities.

BRIDGING CLASSICAL ROBOTICS AND MODERN AI :

IIIT Hyderabad’s autonomous driving team leverages foundational AI models with robust semantic understanding achieving precise localisation and navigation. By augmenting open-source maps with cognitive language markers, the system mimics human-like navigation strategies. This allows the car to adapt to dynamic environments, such as identifying unmarked landmarks like open parking spaces.

This innovative approach integrates classical robotics techniques with advanced AI-driven solutions, striking a balance between traditional methods and cutting-edge technologies.

Effective autonomous navigation depends on three critical components: mapping, localisation, and planning. Traditionally, navigation pipelines separated these tasks, but IIITH’s car adopts a vision-language integration approach, which enhances its ability to process natural language commands like “take a left near the food stall.”

To address the inherent challenges in perception-based navigation, the system employs a differentiable planning module within its neural network framework. This allows real-time error correction and ensures consistent collision-free path planning, even when upstream predictions from perception models are less accurate..

Source : India Today