Autonomous Driving

Benchmark performance should not be directly equated with deployment robustness; laboratory success does not equal real-world robustness.

AIP introduces a controlled intermediate layer between large models and enterprise data, using a combination of a logical rules engine, ontology network, and action sandbox to ensure that AI reasoning steps can be audited, rolled back, and subject to permission controls.

Work that requires deep creativity and fundamental breakthroughs is harder to automate with AI, while standardized, highly repetitive application-layer engineering is more likely to be replaced by AI tools

增强学习领域用'sim-to-real gap'描述在模拟环境训练的机器人策略因现实物理差异而在真实世界失效的现象

自动驾驶从技术演示到真正产品化经历了二三十年

Procedural synthetic data provides robotics and autonomous driving developers with a low-cost, high-precision, and infinitely scalable data acquisition path

In autonomous driving research, a carefully crafted physical object misled a system combining camera and lidar perception; that research did not prove a weakness in any specific robot.

The promotion of autonomous driving features is shifting from "technology-driven" to "experience-driven."

Autonomous driving spent about 20 years and $170 billion in investment after the DARPA Grand Challenge, and its unit economics are still worse than Uber's.

A unified light signal is only part of autonomous driving's external communication; how pedestrians understand this signal and whether other drivers respond correctly still need to be verified through real-world road data and public education.

The bottleneck for autonomous driving has long ceased to be algorithms and now lies in rules.

Behavioral factors such as speeding, drunk driving, and distracted driving require enforcement, education, and culture working together; attributing the elimination of traffic deaths entirely to autonomous driving is a dangerous oversimplification.

In high-risk domains such as healthcare, finance and autonomous driving, the single-minded pursuit of model capability while neglecting verification and calibration is often the root cause of accidents.

Jia Xiaosong believes embodied intelligence is much harder than autonomous driving: contact, friction and deformable objects are difficult to model precisely, the hallucination and speed problems of world models also exist, and the methodology of autonomous driving can be borrowed but one cannot assume it has solved the problems of embodied intelligence.

Generating and manipulating internal visual representations requires substantial computation, posing a real bottleneck for applications requiring real-time response (such as autonomous driving or high-frequency trading).

The industry widely uses the 'long tail theory' to describe the challenge of extreme scenarios in autonomous driving: the first 90% of common scenarios are easy to solve, but the remaining 10% of extreme scenarios require 90% of R&D resources.

In autonomous driving, 99% of road conditions can be handled by the system, and the remaining 1% of long-tail scenarios (such as temporary construction, irregular obstacles, and combinations of extreme weather) often determine the system's safety ceiling.

Taking autonomous driving as an example, perception and decision-making algorithms were close to usable years ago, but what constrains their scale-up is the allocation of accident liability, insurance systems, and local legislation.

In the field of autonomous driving, LiDAR data collection is costly and difficult to annotate, and generating high-quality LiDAR point clouds from multi-view images can lower the barrier to data acquisition.

Autonomous driving is a real-time perception-decision-control system that relies on sensor fusion, BEV perception, and planning and control algorithms, has extremely high requirements for latency and safety, and typically runs on automotive-grade chips.