Oracle Corporation

04/15/2025 | Press release | Distributed by Public on 04/16/2025 22:27

Powering Matic's Autonomous Robotics Revolution with Oracle Cloud Infrastructure

Advancing Home Robotics with True Autonomy
Home robotics has promised to make life easier for decades, yet many people still use bulky, disc-shaped devices that get caught on furniture, tangle with cords, or need constant supervision when they get stuck-wasting more time. After 20 years, it's surprising how little has changed. Matic is reimagining home robotics with fully autonomous, useful, elegant, and affordable robots that see, understand, and handle chores, giving users back time and energy to focus on what truly matters. Creating a truly autonomous indoor robot requires solving some of the hardest problems in computer vision and robotics. Just as Tesla uses vision-only cameras and neural networks to power its Full Self-Driving capabilities, Matic robots rely on cameras and advanced AI to navigate and understand complex indoor environments with unprecedented precision. Behind this innovation is a compute infrastructure that needs to be as cutting-edge as their algorithms-and that's where Oracle Cloud Infrastructure (OCI) comes in.

The Computational Challenge of True Indoor Autonomy
Matic's approach to autonomous indoor robotics differs fundamentally from traditional robots. Instead of relying on basic proximity sensors or expensive LiDAR, Matic uses five RGB cameras to perceive the world, combined with neural networks that process this data on the edge device to enable real-time 3D mapping and navigation. These neural networks don't emerge fully formed-they require extensive training on large amounts of data. Matic's Image-to-Voxel (I2V) neural networks, which create real-time 3D representations of the robot's surroundings, are particularly compute-intensive to train. This is where a reliable high performance AI infrastructure becomes critical and why Matic chose OCI.

Why OCI Powers Matic's AI Training
When training state-of-the-art neural networks, every fraction of a second matters. Slow or inconsistent performance can extend training cycles from days to weeks, directly impacting Matic's ability to iterate quickly and push the boundaries of indoor robotics.

After evaluating several cloud providers, Matic selected OCI for several critical reasons:

Bare Metal Performance Without Compromise OCI's NVIDIA A100 and H100 Tensor Core GPU nodes deliver true bare metal performance without the overhead of virtualization or time-sharing common in many cloud environments. This means training jobs run as if Matic had the hardware in their own data center-no unexpected slowdowns, no mysterious performance degradation.
With OCI's bare metal GPUs, Matic has seen up to 35% faster training times compared to virtualized GPU instances from other providers. This translates to more iterations, better models, and ultimately smarter robots in customers' homes.

Scaling On Demand Matic's training needs fluctuate dramatically. When developing new capabilities or refining existing models, they might suddenly need to scale from tens to hundreds of GPUs. OCI's infrastructure team has consistently provided the required agility.
Recently, when shipping critical features in their neural networks, Matic needed to run a large-scale distributed training job across multiple nodes. Within a day, OCI provisioned the additional dedicated bare metal capacity, allowing Matic to complete a critical training cycle that improved robot models.

Expert Support for AI Workloads ML infrastructure isn't standard cloud computing-it has unique requirements and challenges. OCI's team demonstrates a deep understanding of ML workloads and has been exceptionally responsive, even during off-hours.
One Matic ML engineer recalls sending a Slack message to the OCI team late on a Friday night upon encountering an unexpected bottleneck. OCI's engineers responded quickly and resolved the issue promptly, helping Matic avoid losing a weekend of compute time.

Training Matic's Spatial AI Models on OCI
Matic's Image-to-Voxel neural networks form the foundation of the robots' spatial understanding capabilities. These networks transform raw camera inputs into detailed 3D voxel representations capturing both geometry and semantics of the environment.

Matic implements a multi-stage training pipeline on OCI:

Pre-training on synthetic data: Matic generates millions of synthetic indoor scenes with ground-truth depth and semantics, using OCI's GPU compute clusters for generation and initial training.
Fine-tuning on real-world data: Models are refined using real-world data collected from test environments, enabling generalization to actual homes.
Distillation for edge deployment: The most compute-intensive stage involves distilling these large models into efficient versions that run in real-time on the robots' NVIDIA Jetson hardware.
Each of these stages leverages OCI's infrastructure differently:

For synthetic data generation, Matic utilizes distributed CPU/GPU clusters to create photorealistic renderings of diverse indoor environments.
For primary training, Matic employs multi-node, multi-GPU configurations with NVIDIA A100 GPUs, enabling experimentation with larger batch sizes and more complex architectures.
For model distillation and optimization, NVIDIA A100 GPUs are used for training for edge deployment.
Results: How OCI Powers Real-World Robot Performance
The computational infrastructure provided by OCI can directly translate to tangible improvements in Matic robots' capabilities:

Accurate 3D: Matic's 3D Image-to-Voxel neural networks achieve 1.5cm³ voxel resolution, enabling confident navigation and cleaning even in challenging lighting conditions or dynamic environments.
Semantic Understanding: Robots distinguish between different floor types, identify obstacles, and help understand spatial context (like recognizing kitchens versus living rooms), enabling more intelligent cleaning behaviors.
These capabilities are possible because OCI enables training increasingly sophisticated models without computational constraints. OCI's infrastructure provides computational headroom, meaning the limitations come only from algorithms and data, not from compute resources.

Future Innovations & OCI's Role
As Matic pushes the boundaries of home robotics, OCI remains a key partner in scaling AI training. Upcoming advancements include:

Enhanced object recognition for improved interaction with household items.
Adaptive learning that helps tailor cleaning patterns to specific homes.
Multi-robot coordination for synchronized cleaning in multi-device households.
Each of these features demands more compute power, reinforcing the necessity of OCI's scalable infrastructure.

Why Infrastructure Matters for Robotics Innovation
Creating truly autonomous robots isn't just about algorithms-it's also about having the right infrastructure to develop, train, and refine those algorithms at scale. OCI has proven to be more than just a cloud provider; it has become a true partner in Matic's mission to build the next generation of home robots.

The unseen computational work happening in OCI's data centers can translate to the seamless, reliable experience that Matic customers enjoy in their homes. As Matic continues pushing the boundaries of consumer robotics, OCI remains a foundational element in bringing this vision to life.