Tesla announced on Jun 30 that investigations into a fatal accident involving a Tesla Model S – on Autopilot mode -have been launched. The actual crash took place on May 7 on a highway in Florida, the US, but only later did a statement divulging some details appear on the Tesla website.
(At the bottom of this article: Common misconceptions about the accident, and why it doesn’t necessarily spell doom for the industry.)
According to the statement, the Model S blindsided a large tractor trailer. The latter was driving perpendicular relative to the Tesla, and was midway through a turn. The suite of sensors on the Tesla did not trigger the brakes, resulting in a crash and death of 40-year-old Joshua Brown. The trailer driver was unharmed.
Thus, it’s worth asking whether the car was appropriately equipped to handle the situation. From Tesla’s statement:
Neither Autopilot nor the driver noticed the white side of the tractor trailer against a brightly lit sky, so the brake was not applied.
– Tesla statement “A Tragic Loss”, Jun 30
This invariably points to the onboard optical cameras not being able to distinguish colours perfectly – ergo, had there been a huge pane of clear glass placed across a junction, a Model S would probably run headlong into it, if other sensors were not brought to bear.
That’s where the Tesla’s forward-facing radar is supposed to come in, because a solid object as big as a trailer, invisible or not, would register as an obstacle.
But a tweet from Elon Musk himself hours after the event explains why this failed as well:
— Elon Musk (@elonmusk) June 30, 2016
The long trailer’s perpendicular position and “the high ride height” contributed to it resembling an overhead sign. However, in the same sentence, it’s given that the “bottom of the trailer impact(ed) the windshield of the Model S”.
Following that logic, the bottom of the trailer could not have been more than 1.4 metres (6.5 ft) off the ground – this is the height of a Tesla Model S. This figure is far smaller than the lowest height a sign can be on a US highway – 4.3 metres (14ft). (Source: US Federal Highway Administration’s minimum vertical clearance regulations.)
Judging height, then, could be where the sensors fell short, leading to Autopilot’s decision to try passing under the trailer. Until investigations are completed or further statements are issued, the exact flaw is not clear.
But perhaps Tesla’s rival in the robo-car scene, Google, has the hardware solution on their Google Car prototype: a high-accuracy, laser-based system that, among other merits, is able to map out height, length and width extremely accurately, at ranges up to 100 metres. The white trailer would have registered regardless of colour, and the gap between the road and bottom of the trailer would have been measured as well.
The technology, termed LIDAR for LIght Detection and Ranging, sits atop Google’s prototype cars and churns out 1.3 million readings per second, thus generating a real-time “map” of the cars’ surroundings and obstacles.
Here’s where the bite is. Back in October 2015, Musk said that Google’s use of LIDAR was “unneccesary”. Speaking at a press conference, he said: “I don’t think you need LIDAR. I think you can do this all with passive optical and then with maybe one forward RADAR… I’m not a big fan of LIDAR, I don’t think it makes sense in this context… I think it’s unnecessary.”
“I don’t think you need LIDAR. I think you can do this all with passive optical and then with maybe one forward RADAR… I’m not a big fan of LIDAR, I don’t think it makes sense in this context… I think it’s unnecessary.”
– Elon Musk
To be fair, Musk likely had the eye-raising price tag of a LIDAR system in mind at the time of comment. A single sensor unit used in the earliest Google Car would have cost more than S$100,000, a cost that places such technology out of the average car buyer’s budget.
Google, on the other hand, refused to compromise on a technology it deemed core to a driverless system. The tech giant tackled the cost problem head-on by putting out a job listing for a LIDAR expert, hinting that it’s taking production in-house to lower costs. Their supplier, Velodyne Lidar Inc, also released a scaled-down version which is only slightly less capable, but brings the price down to about a tenth of the original setup.
Lessons learnt? The unfortunate incident shows that it’s better to err on the side of caution in any arena where lives are at stake.
(RELATED: A Tesla Model S was spotted with Lidar sensors in San Francisco, but there is no proof it’s owned or comissioned by Tesla.)
(LATEST: The trailer driver says Brown was watching a movie at the time of the accident – unconfirmed by authorities. A DVD player was also found near the wreckage of the crash.)
Not The End of Assisted or Autonomous Driving
Multiple news reports have covered the facts of the case, and many media outlets have cast doubts on the safety of autonomous cars. Here are some facts of the case that paint a less grim picture:
- The Tesla Model S is NOT an autonomous car, and at no point is the driver supposed to rely fully on the car, even on Autopilot mode, which functions as assisted cruise control and lane-changing functions.
- The same driver had been saved before by the Autopilot mode in a near-miss with a truck. He even uploaded a YouTube video detailing the incident, and expressed gratitude towards the tech for saving his life. Brown was a huge advocate of Tesla and progressive technology – the crash was a truly tragic turn of events.
- Accidents do happen. In fact, of the 130 million miles of Autopilot travel since Oct 2015, this was the first fatal accident. For comparison, in the US, there is one fatality per 94 million miles of regular driving. Using only these metrics, Autopilot can loosely be considered 38% safer.
- With or without AI assistance, an accident could have been unavoidable if the truck cut a light without warning and if both vehicles were travelling at high speed. We will update this page with any news on the investigation.