One perspective that people get wrong about self-driving cars and why Tesla approach won't work
One argument I often hear from Tesla fans is that vision-only should eventually be sufficient to achieve Level 4 autonomy. I have no doubt that a sufficiently advanced vision-only system could eventually outperform humans on many, perhaps even most, driving benchmarks. But I think this misses a bigger issue: the definition of “good enough” tends to change over time.
Look at virtually any safety-critical industry. The safest cars in 1998 were extremely safe by 1998 standards. Yet they could not legally be produced today because crashworthiness, braking, electronic safety systems, pedestrian protection, emissions, and other requirements have continually increased. The same general pattern exists in aviation, construction, medicine, industrial machinery, and other safety-critical fields.
So I don't think the question is simply: “Can vision eventually become better than a human driver?”. I think the more important question is: “Will vision-only continue to satisfy whatever safety and redundancy requirements society considers acceptable 10, 20, or 30 years from now?”. I'm skeptical.
Even if vision-only becomes dramatically safer than human driving, additional independent sensing modalities still provide redundancy and robustness. Lidar, radar, high-definition maps, GNSS/GPS, V2X, and other sources of information can potentially provide independent evidence when the camera system is uncertain, degraded, obstructed, or confronted with an unusual situation.
That doesn't necessarily mean every modality is required in every situation. And it doesn't mean vision-only cannot achieve Level 4 under today's definitions. My argument is about the long-term trajectory of safety standards. If autonomous vehicles become widespread, regulators and the public may eventually demand safety margins far beyond “better than humans.” At that point, a system with multiple independent sources of information may have a fundamental advantage over one relying primarily on a single sensing modality.
This is also why I expect V2X and positioning technologies to become increasingly important. Waymo and other autonomous-driving companies already use forms of mapping and localization, and I wouldn't be surprised if future systems increasingly combine infrastructure/vehicles-derived information (V2X) and GNSS sources as additional layers of redundancy (While Waymo is currently reluctant to adopt V2X & GNSS - I predict it would eventually be "forced" to do so)
There is another assumption I disagree with: that autonomous-driving hardware costs will inevitably fall toward some minimal “commodity” level. I don't think there is necessarily a fixed amount of compute that is simply “enough” for autonomous driving. As compute becomes cheaper, developers can use more of it to improve perception, prediction, planning, simulation, redundancy, uncertainty estimation, edge-case handling, and verification. The same phenomenon happens throughout technology: when a resource becomes cheaper, we often don't simply use less of it - we use substantially more of it to achieve higher performance.