The philosophical difference between Waymo and Tesla driverless taxis

Introduction

I perceive the difference between the driverless taxis of Waymo and of Tesla as being a fundamental and probably, defining, philosophical difference of approach. I suspect it will be “defining” because the “winning” approach now will likely come to dominate the field. Of course, “winning” is a consequence of many economic and social factors as well and not just a measure of a superior technology or a superior philosophy. My assessment is that Waymo approaches autonomous driving as an engineering problem of constructing a sufficiently detailed model of the driving environment, whereas Tesla approaches it as a problem of teaching an artificial intelligence to perceive and navigate the world in much the same way that a human driver does. 

  • Waymo: Builds a machine that knows where it is, knows what is happening around it, and can make algorithmically defined decisions. 
  • Tesla: Builds a machine that learns to see, interpret and respond to the world, without demanding or requiring the world to be extensively mapped and described in advance. The focus is on the intelligent judgement to be made in the prevailing conditions.

Differences of Approach

Waymo tries to build the most accurate representation of the surrounding world as possible using several sensing technologies and even those which humans do not possess. 

  • Cameras for visual information.
  • Lidar for direct three-dimensional distance measurements.
  • Radar for distance and relative velocity, including in conditions where vision is degraded.
  • High-definition maps for precise localisation and advance knowledge of the road environment.

The machine then combines these observations into an operational representation of the world. The correct description of the surrounding world then leads to the selection of the driving decisions to be made.

Tesla starts from a different premise. It assumes that the camera images contain sufficient information to drive, provided the neural networks become sufficiently capable of interpreting them. It therefore puts much greater emphasis on:

  • Camera-based perception.
  • Neural networks trained on enormous quantities of driving data.
  • Learning to recognise situations and predict how they will develop.
  • Improving the driving system through training and software updates.

Rather than explicitly measuring every aspect of the environment using independent instruments, Tesla focuses on getting its AI to make its driving judgements based on the description of  the environment as available from visual observations.

Where the search for excellence resides

This is perhaps the more interesting philosophical distinction. In the Waymo approach, a substantial amount of intelligence is embodied in the engineering of the system itself: the sensors, their calibration, the maps, the perception architecture, the prediction models and the safety constraints. In the Tesla approach, more of the intelligence is expected to emerge from the training of the neural networks. The ambition is to have the network learn the underlying regularities of driving rather than have engineers explicitly specify how every situation should be interpreted and handled.

One could describe these as:

  • Waymo: Excellence in describing the surrounding situation (mapping as well as events) for subsequent selection of actions to be taken
  • Tesla: Excellence of assessment of available information for application of intelligent judgement through learned experience and generalisation.

Neither is free of the other or disqualifies the other. Waymo tries to describe what is in the surroundings now, while Tesla to focus on the action to be taken whatever the now. Waymo’s approach is to reduce uncertainty through additional information and redundancy. If one sensing modality is unreliable, another may provide useful information. Its operational environment is also deliberately constrained. It maps and validates the areas in which it operates. Tesla’s approach is to make the learnt intelligence sufficiently general that it can cope with situations it has not encountered before. It assumes that a sufficiently large and capable neural network can recognise the underlying nature of a situation rather than simply match it against previously encountered examples. This is a fundamental distinction between managing uncertainty through additional data input or by managing it through more general assessment capability. Both approaches have failure modes. Waymo can be confused by situations outside its engineered expectations or operating domain. Tesla can confidently misinterpret something that its visual system has not yet learnt to understand.

For those who can remember the Betamax/VHS wars the winner is a complex result of technology, commercial, and social factors. Scalability could well be decisive. Waymo’s approach requires substantial investment in mapping, specialised hardware, local validation and deployment infrastructure. It can produce a highly capable driverless taxi service, but expanding into a new city is a significant undertaking. Tesla’s ambition is to develop a general driving intelligence that can be deployed across a very large fleet, with relatively little location-specific preparation. If successful, the marginal cost of extending the service to another city could be considerably lower. That is the underlying attraction of Tesla’s approach. It is not merely trying to build a better taxi. It is trying to develop a general-purpose autonomous driving capability.

A personal commentary

I find this particularly interesting when viewed in terms of human cognition. These are two rather different ideas about what it means to understand the world. Waymo’s approach is somewhat analogous to constructing an increasingly accurate scientific model of the external world, using multiple independent instruments to reduce uncertainty. The degrees of freedom needed for then making judgements can be reduced. Tesla’s approach is closer to the proposition that intelligence consists of learning to interpret sensory experience, recognising patterns and developing the capacity to respond appropriately to circumstances. But neither actually tries to reproduce human cognition since the objective is only to drive successfully. Both construct representations of the world that are useful for then taking action. Where they differ again is the manner in which the particular action is decided upon. A human driver does not need to understand the physics of a vehicle, the chemistry of its fuel, the structure of the road or the biology of a pedestrian to drive competently. Driving requires a sufficiently useful model of the world, which can be used by the agent of action to select a particular action.

It brings the distinction between excellence of the driver and excellence of the map into sharp focus. The more detailed the map and the information the less is the capability needed from the driver (the agent of action). The more general the capability of the driver the less detailed the map needs to be. Thus, Tesla is trying to improve the intelligence of the agent while Waymo is trying to improve the information available to the agent. Tesla is betting that intelligence can compensate for incomplete information. The driver must learn to extract meaning from relatively ordinary sensory inputs, make inferences, anticipate events and respond appropriately. This is broadly analogous to human cognition. Waymo is betting that better information can compensate for limitations in the intelligence of the driver. It invests in detailed environmental measurement, mapping and engineered representations, thereby reducing the amount of inference required of the driving system. Of course, Waymo’s driver is not stupid, and Tesla’s driver is not merely a human imitation. But the contrast in strategic emphasis is real.

It is also a philosophic distinction between being general and being specific. Tesla’s ambition seems to be to develop a driving intelligence that can encounter an unfamiliar environment and work out how to navigate it. Its intelligence thus is embedded in, and travels with, the vehicle. Waymo’s approach is more dependent on preparing and validating the environment in which the driver operates. Its environmental knowledge is partly externalised rather than residing entirely within the driving intelligence. But one consequence is that if the extra information is not available then the driver is paralysed. This gives Tesla a potential scaling advantage if its approach works. It also gives Waymo a potentially significant reliability advantage in environments it has thoroughly characterised.

In reality, of course, both the map and the agent of action are  needed. It is a matter of emphasis even if the philosophical distance is quite fundamental. Intelligence must be sufficiently capable of extracting sufficient understanding from the available information. At the same time information must be sufficient to allow the available intelligence to act successfully. These are not mutually exclusive engineering strategies, but they represent different answers to a fundamental question about the relationship between cognition and the world.

There is also an interesting evolutionary analogy. Humans did not evolve by acquiring lidar, radar and increasingly elaborate maps of their surroundings. We evolved sensory and cognitive capabilities that allowed us to make useful models of our surroundings from incomplete information. Drowning a brain in sensory data it was not evolved to handle is nearly always counter-productive. A human brain cannot interpret direct radar signals and has to handle them via vision. So in a car equipped with a variety of senses, its driving brain will need to be able to interpret the signals directly. A plethora of different sensory data  may increase the data available but will, always, also increase the conflict between interpretations. For example providing a human brain with ultraviolet signals is of no value at all. Radar signals can only be detected by humans though their visual sensory equipment and cannot be interpreted by human brains. So if a vehicle gets conflicting signals, say by vision and by radar, deciding which to reject and which to accept can also lead to paralysis. Tesla is, in this respect, pursuing something closer to an evolutionary approach to intelligence. Waymo is pursuing something closer to an engineering approach to environmental certainty.

Ultimately the winner is not an indicator of a superior underlying philosophy. Economics, safety regulation, deployment costs and public acceptance will finally determine the outcome.  But right now a succinct description of the strategic difference, and potentially of the defining technological competition in autonomous driving is that:

Waymo is seeking excellence of the map.

Tesla is seeking excellence of the driver. 


 

Tags: , , , , , , , , , ,

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.