Motive Builds Driver Safety AI on Anyscale

Motive is an AI-first platform for physical operations, providing fleet management, driver safety, and asset tracking technology to over 120,000 customers across transportation and logistics. At the core of their product are computer vision models trained on video from more than one million connected dashcams deployed across North America and the UK. Those models power customer-facing safety features including stop sign violation detection, lane departure warnings, driver fatigue alerts, and distraction monitoring, with each validated model version rolling out to hundreds of thousands of cameras on the road.

Behind that product is a continuous model development cycle: training perception models on large video datasets, running extensive regression testing across thousands of video events, and shipping new model versions within tight release windows. For years, the infrastructure supporting that cycle could not keep pace. With Ray on Anyscale, Motive can train and test models using billions of video frames on a single, scalable platform. 

LinkChallenges

Continuously improving AI safety using thousands of cameras demands infrastructure that can keep up with the pace of the necessary training and testing. As Motive's AI team and its roster of production models grew, resource contention, sequential testing bottlenecks, and the operational overhead of managing siloed environments were costing the team cycles that could be used to improve their customer’s driver safety and productivity.

Three challenges stood in the way:

• Slow training and sequential testing were slowing down the model development cycle. Before Anyscale, Motive's teams were in a constant manual battle for GPU compute. Training ran on hand-provisioned EC2 instances without distributed processing. Testing compounded the problem further: running thousands of dashcam video events through a multi-stage pipeline covering frame preprocessing, perception model inference, temporal classification, event validation, and database writes ran sequentially, stretching a full regression test across two to three days which routinely pushed rollouts past their deadline.

• No unified platform meant every new model project started from scratch. Motive's workflows were split across disconnected environments for data processing, training, and inference, with engineers across the US, Pakistan, and India manually coordinating GPU access in Slack just to start a job. Every new model project required reinventing the operational scaffolding, and a small platform team was left supporting an ever-growing number of researchers and engineers with no scalable way to do so.

Growing cloud costs compounded with underutilized GPUs. As Motive made the shift from deploying workloads on office servers to Amazon EC2, the team unlocked scale but created a new cost efficiency problem. Every workload ran on on-demand instances with no spot coverage, GPU utilization across training jobs was far below what the hardware was capable of, and there was no mechanism to right-size compute across projects or prevent engineers from spinning up instances that sat underutilized. As the number of models and engineers grew, so did the compute waste.

"Placing workloads was a manual exercise where engineers across the team negotiated GPU access in Slack. It got the job done, but it wasn't a system anyone could scale on. "

Tim Cheng | Engineering Manager, Motive

LinkThe Solution

Motive ran proof-of-concept experiments across both training and testing use cases in early 2025, confirming that Anyscale could handle both efficiently before committing to a broader migration. What began with two pilot models and a handful of engineers has since grown to over thirty engineers submitting jobs regularly across ten active projects, with more than twelve production models now trained on the platform.

With Anyscale, Motive is able to:

• Accelerate training and testing with distributed Ray Train and Ray Data pipelines, with Anyscale Agent Skills lowering the barrier to entry. By migrating from single-node EC2 to Ray Train, the team fans out training jobs across sixteen or more L4 GPU nodes simultaneously with no manual provisioning or coordination required. The model testing pipeline has been rebuilt as a Ray Data workflow that streams preprocessing, perception model inference, temporal logic, and event classification in parallel, with CPU and GPU resources scaled independently for each stage. Engineers can also submit, monitor, and debug jobs using Anyscale Agent Skills without needing deep familiarity with the underlying infrastructure, compressing the time from experiment to result across every team. 

• Support a growing team of thirty-plus engineers with self-service access to compute across ten projects. Rather than negotiating GPU time over Slack or waiting on a platform team to provision resources, engineers submit jobs through consistent tooling that works the same way across every model project. The Anyscale Workload Scheduler manages priority-aware queuing across all active workloads, giving teams visibility into job status without requiring manual intervention from the platform team. Even Motive's Director of AI was submitting training jobs on Anyscale within his first week of onboarding.

Reduce infrastructure costs by over $180,000 per year through smarter compute utilization. Moving off physical office servers eliminated the fixed cost of on-premises hardware that could not scale with demand. On the cloud side, the team is actively leveraging on-demand instances, while optimizing GPU utilization for higher throughput per job through improved batch sizes, data loading configurations, and precision settings. Together, these efforts are projected to save over $180,000 annually in compute costs, reduce infrastructure headcount growth by two FTEs, and deliver a fifteen percent lift in developer productivity.

"Running data processing and training in a single compute pool has simplified everything. We can mix CPU and GPU workloads in the same pipeline while scaling each stage independently, making our end-to-end workflows faster and more efficient. "

Tim Cheng | Engineering Manager, Motive

LinkFaster training and testing across every stage of the pipeline

Anyscale serves as the substrate for Motive's full model development lifecycle, covering the road-facing models that detect vehicles, lane markings, and traffic signs and the driver-facing models that classify fatigue, distraction, and unsafe behavior. Both model families process large volumes of dashcam video through a multi-stage pipeline, and both now run on the same platform without separate infrastructure configurations for each.

On the training side, Motive uses Ray Train with PyTorch and Weights and Biases for experiment logging. Engineers package their model code against shared internal base images, submit jobs through a standardized template, and see results in consistent dashboards regardless of which model they are working on. Key production models like the Vision Based Collision model now train in six hours using 64 L4 GPUs, down from seven days on single-node EC2 instances, 28x faster. The team has grown from two pilot models to more than twelve in production. Paired with Anyscale Agent Skills, engineers can submit, monitor, and debug training jobs without needing deep infrastructure expertise, reducing the time from experiment to production even further.

The test infrastructure required rebuilding from the ground up, since Motive's existing platform ran thousands of video events sequentially through preprocessing, model inference, temporal classification, and event validation before writing results to a database. Rebuilt on Ray Data, testing now runs as an end-to-end batch inference pipeline with heterogeneous compute: CPUs handle preprocessing, GPUs run model inference, and each stage scales independently. A full regression run that previously took three days now completes in fifteen minutes, 288x faster. Motive previously iterated on one to two model versions per release cycle and now runs ten to twenty iterations in the same window, driving drastic accuracy improvements across every production model.

"With Anyscale, we can distribute training and testing across the cluster using the same frameworks our teams already know, like PyTorch. Paired with Anyscale Agent Skills, teams are running the end-to-end lifecycle in hours instead of close to a full week. "

Tim Cheng | Engineering Manager, Motive

LinkGiving every engineer access to compute without bottlenecking the platform team

When Tim's team first started using Anyscale, the active user base numbered three to five engineers. Within months, that number grew to over thirty engineers submitting jobs regularly across ten projects. The standardized templates and consistent tooling that the AI infrastructure team built on top of Anyscale made it possible for model training teams to adopt the platform without needing deep infrastructure expertise.

With Anyscale, engineers who previously had to wait hours or negotiate machine access over Slack could submit a job and see results within a session. Motive's Director of AI was among the earliest adopters, submitting training jobs on Anyscale within his first week at the company. The team manages quotas and job queues internally today and is actively working toward integrating the Anyscale Workload Scheduler across all workloads to bring more structure to how compute is allocated between research and production priorities, giving every engineer visibility into job status without escalating to the platform team.

"People don't want to disrupt their day-to-day habits, but now the incentives are we can run jobs much faster and are easier to set up, and there's no resource contention anymore. Even our director of AI was submitting a job on Anyscale in his first week of onboarding."

Tim Cheng | Engineering Manager, Motive

LinkEliminating compute waste spend across a dozen model projects

Running training on physical office servers required dedicated hardware that could not flex with demand and concentrated expensive GPU resources into a fixed pool that every engineer competed for. While the move to cloud compute solved much of the scaling problem, it introduced a new one. With every workload running on on-demand instances and no way to right-size compute across projects, GPU utilization stayed low and idle instances accumulated unnoticed as the team grew.

Motive currently runs all workloads on on-demand instances, with ongoing optimization work with Anyscale targeting higher throughput per job through better batch sizes, data loading configurations, and precision settings. These efforts are projected to save over $180,000 annually in compute costs, reduce infrastructure headcount growth by two FTEs, and deliver a fifteen percent lift in developer productivity.

"Anyscale provides our team with a unified compute pool where we can intelligently schedule our training and inference AI workloads, and we can mix spot instances into that same pool to bring cloud costs down without changing how anyone submits jobs."

Tim Cheng | Engineering Manager, Motive

LinkWhat's Next

Looking ahead, Motive is expanding its AI platform in several directions: building a Driver Foundation Model trained on billions of dashcam data points currently sitting unused in their data lake, migrating cloud inference workloads to a more flexible serving architecture, and pushing toward larger and more complex models that will require training across many more nodes than their current workloads demand. As the fleet grows and the models become more sophisticated, the infrastructure investments the team has made will need to scale with them.

"Our near-term roadmap includes expanding our foundation model training and migrating online inference to Ray Serve. We have billions of data points we have barely touched, and Anyscale is the platform we expect to use to unlock them. "

Tim Cheng | Engineering Manager, Motive