We deployed 100 reinforcement studying (RL)-controlled automobiles into rush-hour freeway visitors to clean congestion and cut back gas consumption for everybody. Our purpose is to deal with “stop-and-go” waves, these irritating slowdowns and speedups that normally haven’t any clear trigger however result in congestion and vital vitality waste. To coach environment friendly flow-smoothing controllers, we constructed quick, data-driven simulations that RL brokers work together with, studying to maximise vitality effectivity whereas sustaining throughput and working safely round human drivers.
Total, a small proportion of well-controlled autonomous autos (AVs) is sufficient to considerably enhance visitors circulate and gas effectivity for all drivers on the highway. Furthermore, the skilled controllers are designed to be deployable on most fashionable autos, working in a decentralized method and counting on customary radar sensors. In our newest paper, we discover the challenges of deploying RL controllers on a large-scale, from simulation to the sector, throughout this 100-car experiment.
The challenges of phantom jams
A stop-and-go wave shifting backwards by means of freeway visitors.
When you drive, you’ve certainly skilled the frustration of stop-and-go waves, these seemingly inexplicable visitors slowdowns that seem out of nowhere after which out of the blue clear up. These waves are sometimes brought on by small fluctuations in our driving conduct that get amplified by means of the circulate of visitors. We naturally modify our velocity based mostly on the automobile in entrance of us. If the hole opens, we velocity as much as sustain. In the event that they brake, we additionally decelerate. However on account of our nonzero response time, we’d brake only a bit tougher than the automobile in entrance. The following driver behind us does the identical, and this retains amplifying. Over time, what began as an insignificant slowdown turns right into a full cease additional again in visitors. These waves transfer backward by means of the visitors stream, resulting in vital drops in vitality effectivity on account of frequent accelerations, accompanied by elevated CO2 emissions and accident danger.
And this isn’t an remoted phenomenon! These waves are ubiquitous on busy roads when the visitors density exceeds a crucial threshold. So how can we deal with this downside? Conventional approaches like ramp metering and variable velocity limits try to handle visitors circulate, however they usually require expensive infrastructure and centralized coordination. A extra scalable strategy is to make use of AVs, which might dynamically modify their driving conduct in real-time. Nevertheless, merely inserting AVs amongst human drivers isn’t sufficient: they need to additionally drive in a wiser approach that makes visitors higher for everybody, which is the place RL is available in.
Elementary diagram of visitors circulate. The variety of automobiles on the highway (density) impacts how a lot visitors is shifting ahead (circulate). At low density, including extra automobiles will increase circulate as a result of extra autos can go by means of. However past a crucial threshold, automobiles begin blocking one another, resulting in congestion, the place including extra automobiles really slows down total motion.
Reinforcement studying for wave-smoothing AVs
RL is a strong management strategy the place an agent learns to maximise a reward sign by means of interactions with an setting. The agent collects expertise by means of trial and error, learns from its errors, and improves over time. In our case, the setting is a mixed-autonomy visitors state of affairs, the place AVs study driving methods to dampen stop-and-go waves and cut back gas consumption for each themselves and close by human-driven autos.
Coaching these RL brokers requires quick simulations with sensible visitors dynamics that may replicate freeway stop-and-go conduct. To attain this, we leveraged experimental knowledge collected on Interstate 24 (I-24) close to Nashville, Tennessee, and used it to construct simulations the place autos replay freeway trajectories, creating unstable visitors that AVs driving behind them study to clean out.
Simulation replaying a freeway trajectory that reveals a number of stop-and-go waves.
We designed the AVs with deployment in thoughts, guaranteeing that they will function utilizing solely primary sensor details about themselves and the automobile in entrance. The observations include the AV’s velocity, the velocity of the main automobile, and the house hole between them. Given these inputs, the RL agent then prescribes both an instantaneous acceleration or a desired velocity for the AV. The important thing benefit of utilizing solely these native measurements is that the RL controllers will be deployed on most fashionable autos in a decentralized approach, with out requiring extra infrastructure.
Reward design
Essentially the most difficult half is designing a reward perform that, when maximized, aligns with the totally different targets that we want the AVs to realize:
- Wave smoothing: Scale back stop-and-go oscillations.
- Power effectivity: Decrease gas consumption for all autos, not simply AVs.
- Security: Guarantee cheap following distances and keep away from abrupt braking.
- Driving consolation: Keep away from aggressive accelerations and decelerations.
- Adherence to human driving norms: Guarantee a “regular” driving conduct that doesn’t make surrounding drivers uncomfortable.
Balancing these targets collectively is troublesome, as appropriate coefficients for every time period have to be discovered. For example, if minimizing gas consumption dominates the reward, RL AVs study to come back to a cease in the course of the freeway as a result of that’s vitality optimum. To forestall this, we launched dynamic minimal and most hole thresholds to make sure protected and cheap conduct whereas optimizing gas effectivity. We additionally penalized the gas consumption of human-driven autos behind the AV to discourage it from studying a egocentric conduct that optimizes vitality financial savings for the AV on the expense of surrounding visitors. Total, we goal to strike a stability between vitality financial savings and having an affordable and protected driving conduct.
Simulation outcomes
Illustration of the dynamic minimal and most hole thresholds, inside which the AV can function freely to clean visitors as effectively as attainable.
The standard conduct discovered by the AVs is to keep up barely bigger gaps than human drivers, permitting them to soak up upcoming, presumably abrupt, visitors slowdowns extra successfully. In simulation, this strategy resulted in vital gas financial savings of as much as 20% throughout all highway customers in essentially the most congested eventualities, with fewer than 5% of AVs on the highway. And these AVs don’t must be particular autos! They’ll merely be customary client automobiles outfitted with a wise adaptive cruise management (ACC), which is what we examined at scale.
Smoothing conduct of RL AVs. Purple: a human trajectory from the dataset. Blue: successive AVs within the platoon, the place AV 1 is the closest behind the human trajectory. There may be sometimes between 20 and 25 human autos between AVs. Every AV doesn’t decelerate as a lot or speed up as quick as its chief, resulting in reducing wave amplitude over time and thus vitality financial savings.
100 AV area take a look at: deploying RL at scale


Our 100 automobiles parked at our operational middle in the course of the experiment week.
Given the promising simulation outcomes, the pure subsequent step was to bridge the hole from simulation to the freeway. We took the skilled RL controllers and deployed them on 100 autos on the I-24 throughout peak visitors hours over a number of days. This massive-scale experiment, which we known as the MegaVanderTest, is the biggest mixed-autonomy traffic-smoothing experiment ever carried out.
Earlier than deploying RL controllers within the area, we skilled and evaluated them extensively in simulation and validated them on the {hardware}. Total, the steps in direction of deployment concerned:
- Coaching in data-driven simulations: We used freeway visitors knowledge from I-24 to create a coaching setting with sensible wave dynamics, then validate the skilled agent’s efficiency and robustness in a wide range of new visitors eventualities.
- Deployment on {hardware}: After being validated in robotics software program, the skilled controller is uploaded onto the automobile and is ready to management the set velocity of the automobile. We function by means of the automobile’s on-board cruise management, which acts as a lower-level security controller.
- Modular management framework: One key problem in the course of the take a look at was not accessing the main automobile info sensors. To beat this, the RL controller was built-in right into a hierarchical system, the MegaController, which mixes a velocity planner information that accounts for downstream visitors situations, with the RL controller as the ultimate determination maker.
- Validation on {hardware}: The RL brokers had been designed to function in an setting the place most autos had been human-driven, requiring sturdy insurance policies that adapt to unpredictable conduct. We confirm this by driving the RL-controlled autos on the highway below cautious human supervision, making adjustments to the management based mostly on suggestions.

Every of the 100 automobiles is related to a Raspberry Pi, on which the RL controller (a small neural community) is deployed.

The RL controller instantly controls the onboard adaptive cruise management (ACC) system, setting its velocity and desired following distance.
As soon as validated, the RL controllers had been deployed on 100 automobiles and pushed on I-24 throughout morning rush hour. Surrounding visitors was unaware of the experiment, guaranteeing unbiased driver conduct. Information was collected in the course of the experiment from dozens of overhead cameras positioned alongside the freeway, which led to the extraction of tens of millions of particular person automobile trajectories by means of a pc imaginative and prescient pipeline. Metrics computed on these trajectories point out a development of lowered gas consumption round AVs, as anticipated from simulation outcomes and former smaller validation deployments. For example, we will observe that the nearer individuals are driving behind our AVs, the much less gas they seem to devour on common (which is calculated utilizing a calibrated vitality mannequin):
Common gas consumption as a perform of distance behind the closest engaged RL-controlled AV within the downstream visitors. As human drivers get additional away behind AVs, their common gas consumption will increase.
One other method to measure the impression is to measure the variance of the speeds and accelerations: the decrease the variance, the much less amplitude the waves ought to have, which is what we observe from the sector take a look at knowledge. Total, though getting exact measurements from a considerable amount of digicam video knowledge is difficult, we observe a development of 15 to twenty% of vitality financial savings round our managed automobiles.
Information factors from all autos on the freeway over a single day of the experiment, plotted in speed-acceleration house. The cluster to the left of the purple line represents congestion, whereas the one on the correct corresponds to free circulate. We observe that the congestion cluster is smaller when AVs are current, as measured by computing the realm of a tender convex envelope or by becoming a Gaussian kernel.
Remaining ideas
The 100-car area operational take a look at was decentralized, with no express cooperation or communication between AVs, reflective of present autonomy deployment, and bringing us one step nearer to smoother, extra energy-efficient highways. But, there may be nonetheless huge potential for enchancment. Scaling up simulations to be quicker and extra correct with higher human-driving fashions is essential for bridging the simulation-to-reality hole. Equipping AVs with extra visitors knowledge, whether or not by means of superior sensors or centralized planning, may additional enhance the efficiency of the controllers. For example, whereas multi-agent RL is promising for bettering cooperative management methods, it stays an open query how enabling express communication between AVs over 5G networks may additional enhance stability and additional mitigate stop-and-go waves. Crucially, our controllers combine seamlessly with current adaptive cruise management (ACC) techniques, making area deployment possible at scale. The extra autos outfitted with good traffic-smoothing management, the less waves we’ll see on our roads, that means much less air pollution and gas financial savings for everybody!
Many contributors took half in making the MegaVanderTest occur! The total listing is out there on the CIRCLES mission web page, together with extra particulars concerning the mission.
Learn extra: [paper]