DigestAI news desk
hot Microsoft AI CEO warns Anthropic's model rights approach risks alignment failures new Legal LLMs' Hallucinations Should Be Evaluated as Warrant Failures hot TypeSafe AI unveils Jev, a frontier model up to 400× cheaper and 200× faster new Benchmarking LLMs for Key-Value Extraction in Noisy OCR Documents hot Google DeepMind launches Gemini 3.8 Live and 3.8 Live Extended Thinking models new Study Shows Relation Facts Trigger Earlier Than Entity Facts in Language Models hot JiT-DDT Architecture Trains Text-to-Image Models 3.6x Faster new LLM-Enhanced Model Improves Extubation Failure Prediction Using Therapy Notes hot OpenAI launches Sponsored Agents and AI ad tools in ChatGPT new Study maps four-stage pipeline for LLM math word problems, isolates failure point hot BITCOS layout cuts ternary LLM storage below 1.58 bits per weight new FairCompressAgent: An Agentic Framework for Fairness-Aware Model Compression hot Anthropic merges Cowork and chat into unified 'One Claude' platform new NeMo Data Designer Offers Flexible Framework for Multimodal Synthetic Data hot Apple Unveils macOS 27 Golden Gate with Built‑In Generative AI and Intel‑Free Design new Nvidia’s Inference Opportunity: Beyond Training
Research updated

Compact Multimodal Imitation Policy Drives Hours Collision-Free in CARLA

The paper presents a compact multimodal driving policy trained via behavioral cloning on offline expert demonstrations and evaluated in the CARLA simulator. The policy ingests five‑frame histories of RGB images, LiDAR point clouds, vehicle telemetry, and lane waypoints, outputting throttle, brake and steering commands at 20 Hz.

1 source primary source

Key points

  • 1.36 M‑parameter multimodal policy trained on 3.3 h of expert data (236,882 windows)
  • Achieves hours of collision‑free autonomous driving on both seen and unseen CARLA routes
  • Public release includes code, checkpoint, ONNX model, data sample, and audit documentation

Training used 236,882 short windows—about 3.3 hours of driving collected from 448 separate captures—resulting in a 1.36 million‑parameter network. In closed‑loop tests the model navigated both training routes and previously unseen held‑out routes for multiple hours without collisions, and it transferred qualitatively to a different CARLA town with distinct road geometry. The authors also note occasional recovery from large trajectory deviations, though systematic recovery was not formally evaluated.

All code, the trained checkpoint, an ONNX export, sample data, and an evidence audit are released publicly, enabling reproducibility and further research on offline imitation learning for autonomous driving.

Read the original at arXiv cs.AI · by Jordy Kieto primary source Open source ↗

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us.

Comments

via GitHub Discussions

More in Research

All →

Related stories