NVIDIA DRIVE

DeepDrive (Self-driving Simulator & Research Tool) 

• What it is:
  An open-source driving simulator and Python API (via gym) that enables experimentation with self-driving agents—ideal for training and evaluating vision-based driving models.

• How it works:
  Built on Unreal Engine + OpenGL, it supports reinforcement learning agents using TensorFlow, with integrations via Docker. GPU performance (e.g., GTX 980) impacts frame rate: typically ~20 fps with eight 512×512 sensors, or ~50 fps with a single sensor.

• Getting started:

  1. Pull the Docker container: deepdriveio/ue4-deepdrive-deps:latest

  2. Bind-mount your Unreal project and Deepdrive source inside Docker.

  3. Install CUDA/cuDNN and NVIDIA drivers compatible with your system.

• Use cases:
  Great for experimenting with imitation learning, DAgger-style fine-tuning, baseline agents, reward functions, and custom sensors in simulation.

2. NVIDIA DRIVE / DRIVE AGX (Full-stack Autonomous Vehicle Platform)

• What it is:
  NVIDIA's flagship platform for automotive AI — combining hardware (DRIVE AGX, Orin, Thor SOC), software (DriveOS, DriveWorks, DRIVE AV), and tools (Omniverse/Cosmos for simulation, DGX for training).

• What it does:

  • In-vehicle compute: DRIVE AGX runs deep neural nets to process camera, lidar, radar, ultrasonic inputs in real time — up to 2,000 TOPS on Thor, 254 TOPS on Orin.

  • Software stack: Includes perception, localization, planning, control — supports end-to-end planning via deep learning DNNs.

  • Simulation & safety: NVIDIA Omniverse and Cosmos simulate diverse scenarios; DRIVE AGX Hyperion integrates sensors; Halos adds safety guardrails.

  • Mapping: DRIVE Mapping, powered by the DeepMap acquisition, enables HD mapping and localization using camera + radar.

• Development flow:

1. Model training on DGX clusters or cloud (with DRIVE DNNs, NGC, foundation models).

2. Simulated validation via Omniverse/Cosmos.

3. Vehicle deployment on DRIVE AGX with safety-certified middleware.

4. Continuous mapping & fleet feedback through DRIVE Mapping.

End-to-End Ecosystem

• Software stack:
  DriveOS (runs on Linux or QNX Safety OS), DriveWorks for sensor fusion, Omniverse/DRIVE Sim for high-fidelity simulation and validation.

• Safety & certification:
  DRIVE Orin and Hyperion platforms have achieved ISO 26262 ASIL‑D compliance and TÜV safety/cybersecurity certification.

• Industry adoption:

  • Orin used by Toyota, GM, Volvo, Nuro, Gatik, Uber Freight, Torc, etc.

  • Thor already adopted by BYD, XPENG, Li Auto, ZEEKR, Nuro, Plus, Waabi, WeRide.