Preferred Networks Cracks Large-Scale Robot Sim

Preferred Networks (PFN), one of Japan's most valuable AI startups, has announced a major breakthrough in large-scale robot simulation that enables thousands of virtual robots to train simultaneously in complex environments. The advance could dramatically accelerate the development of real-world robotic systems by reducing training time from weeks to hours.

The Tokyo-based company, valued at over $3.5 billion, revealed that its new simulation platform can run up to 10,000 parallel robot instances on a single GPU cluster — a roughly 100x improvement over conventional approaches used by competitors like NVIDIA's Isaac Sim or Google DeepMind's MuJoCo-based pipelines.

Key Takeaways

• Scale: PFN's platform supports up to 10,000 parallel robot simulations on a single GPU cluster
• Speed: Training tasks that previously took 2-3 weeks now complete in under 8 hours
• Fidelity: Physics accuracy reportedly matches real-world behavior with less than 2% deviation in benchmark tests
• Transfer: Sim-to-real transfer rates exceed 90% on manipulation tasks, significantly above the industry average of 60-70%
• Cost: Estimated 80% reduction in compute costs compared to existing large-scale simulation frameworks
• Compatibility: The platform integrates with standard robotics middleware including ROS 2 and common URDF robot models

How PFN Achieved the 100x Simulation Speedup

The breakthrough centers on a novel GPU-native physics engine that PFN developed entirely in-house. Unlike traditional robot simulators that rely on CPU-based physics calculations and then offload rendering to GPUs, PFN's approach keeps the entire simulation pipeline — physics, collision detection, sensor rendering, and policy inference — on the GPU.

This architecture eliminates the costly data transfers between CPU and GPU memory that typically bottleneck large-scale simulations. PFN's engineers reportedly rewrote core physics algorithms using custom CUDA kernels optimized for massively parallel execution.

The company also introduced a proprietary scene graph compression technique that reduces the memory footprint of each simulation instance by approximately 95%. This means a GPU cluster with 8 NVIDIA H100 GPUs can host 10,000 robot instances simultaneously, compared to roughly 100-200 instances on comparable hardware using NVIDIA's Isaac Sim.

Sim-to-Real Transfer Breaks the 90% Barrier

Perhaps the most significant aspect of PFN's announcement is the sim-to-real transfer rate — the degree to which behaviors learned in simulation successfully translate to physical robots. PFN reports achieving greater than 90% transfer success on complex manipulation tasks, including picking irregularly shaped objects, assembling multi-part components, and navigating cluttered warehouse environments.

This figure represents a substantial leap over the industry norm. Most robotics labs report sim-to-real transfer rates between 60% and 70% for manipulation tasks, with the gap typically attributed to imperfect physics modeling, sensor noise discrepancies, and actuator dynamics that simulators fail to capture accurately.

PFN attributes its high transfer rates to several innovations:

• Domain randomization at scale: With 10,000 parallel instances, the system can explore vastly more environmental variations during training
• Learned physics residuals: A neural network corrects systematic errors in the physics engine by learning from small amounts of real-world data
• Adaptive sensor noise models: Simulated camera and force-torque sensor outputs are calibrated against real sensor data using automated pipelines
• Contact dynamics refinement: A specialized module models soft-body deformation and friction with higher fidelity than standard rigid-body simulators

Why This Matters for the Global Robotics Industry

The robotics industry has long recognized simulation as the key to scaling robot learning. Training robots in the real world is expensive, slow, and potentially dangerous. Companies like NVIDIA, Google DeepMind, and Boston Dynamics have invested heavily in simulation infrastructure, but progress has been incremental.

PFN's breakthrough arrives at a critical moment. The global robotics market is projected to reach $218 billion by 2030, according to Fortune Business Insights, with warehouse automation, manufacturing, and healthcare robotics driving much of that growth. However, programming robots for new tasks remains painfully slow — a single pick-and-place task can take weeks of engineering effort.

Large-scale simulation fundamentally changes this equation. If a robot can learn a new task in 8 hours of simulated training and then execute it in the real world with 90%+ reliability, the economics of deploying robots in new environments shift dramatically. Small and mid-sized manufacturers that previously couldn't justify the integration costs could become viable customers.

PFN has already demonstrated this capability with its partner Fanuc, one of the world's largest industrial robot manufacturers. The two companies have reportedly used the new simulation platform to train Fanuc robots for 12 different bin-picking scenarios in a single weekend — work that previously required months of manual programming and real-world testing.

How PFN Compares to Western Competitors

PFN's announcement places it in direct competition with several major Western players in the robot simulation space. Here's how the landscape looks:

NVIDIA Isaac Sim remains the most widely adopted commercial platform, with strong GPU acceleration and photorealistic rendering. However, its parallel simulation capacity typically maxes out at several hundred instances per GPU cluster, and its physics engine still relies on PhysX, which was originally designed for gaming rather than robotics precision.

Google DeepMind has used MuJoCo extensively for reinforcement learning research, achieving impressive results on locomotion and manipulation benchmarks. MuJoCo is lightweight and fast for individual simulations but lacks native support for the kind of massively parallel execution PFN demonstrates.

Tesla's Optimus team has built internal simulation tools for humanoid robot training, but details remain scarce, and the technology isn't available to the broader robotics community.

PFN's platform differentiates itself on raw parallel scale and sim-to-real transfer quality. However, it remains to be seen whether the company will open-source or commercialize the platform for external developers — a decision that could significantly influence its industry impact.

PFN's Broader AI Strategy and Track Record

Preferred Networks has been a quiet powerhouse in Japanese AI since its founding in 2014. The company originally gained recognition for developing Chainer, one of the first define-by-run deep learning frameworks that influenced the design of PyTorch. Although PFN eventually migrated to PyTorch itself, the Chainer project demonstrated the company's ability to produce foundational AI infrastructure.

The company maintains deep partnerships with Japanese industrial giants including Toyota, Fanuc, and NTT. Its AI applications span autonomous driving, drug discovery, materials science, and robotics — a breadth unusual for a startup, even one valued at $3.5 billion.

In robotics specifically, PFN gained international attention in 2020 when it demonstrated a home cleaning robot that could autonomously tidy a messy room. While the demo appeared modest compared to the flashy humanoid robots from Tesla or Figure AI, it showcased PFN's focus on practical, deployable AI rather than headline-grabbing prototypes.

This latest simulation breakthrough builds on that philosophy. Rather than building a humanoid robot, PFN is building the infrastructure that makes all robots smarter, faster.

What This Means for Developers and Businesses

For robotics developers, PFN's breakthrough signals an important shift. Simulation-first development — where robots learn entirely in virtual environments before touching the real world — is becoming viable at production scale, not just in research labs.

Businesses evaluating robotic automation should watch for several downstream effects:

• Faster deployment timelines: Weeks instead of months to configure robots for new tasks
• Lower integration costs: Reduced need for expensive real-world testing and manual programming
• Greater flexibility: Robots that can be retrained for new products or environments overnight
• Improved reliability: Higher sim-to-real transfer means fewer failures and less downtime in production

For the broader AI community, PFN's work also highlights the growing importance of embodied AI — training AI systems that interact with the physical world, not just process text or images. As large language models face diminishing returns on benchmark performance, embodied AI represents a frontier where massive compute investments can unlock tangible real-world value.

Looking Ahead: What Comes Next

PFN has indicated it will publish a detailed technical paper on the simulation platform in the coming months, likely at a major robotics conference such as ICRA 2025 or CoRL. The company is also reportedly exploring a limited commercial release of the platform, potentially through cloud-based access for enterprise customers.

Several questions remain. Will PFN open the platform to the broader robotics community, or keep it as a competitive advantage for its industrial partners? Can the 90%+ sim-to-real transfer rates hold up across a wider variety of robot morphologies and task types? And will Western competitors — particularly NVIDIA, which has significant resources and market presence — respond with comparable capabilities?

What is clear is that the era of large-scale robot simulation has arrived. PFN's breakthrough demonstrates that the computational tools to train thousands of robots simultaneously now exist, and the gap between simulated and real-world performance is narrowing fast. For an industry that has long promised transformative automation but struggled with the slow pace of robot programming, this could be the infrastructure breakthrough that finally accelerates deployment at scale.