SIGGRAPH 2025 Special Address — AI, Simulation, and the Dawn of Graphics 3.0

I’m writing this report in the immediate aftermath of our SIGGRAPH 2025 special address, presented by NVIDIA. In that talk I shared the stage with my colleagues Sanja Fidler, Aaron Lefohn, and Ming-Yu, and together we painted a picture of where graphics and AI are headed. Our message was simple, urgent, and optimistic: AI is reshaping what’s possible in graphics, and graphics is reshaping what’s possible in AI. In this piece I’ll summarize the key ideas we presented, expand on their implications, and map out what I see as the research and industry priorities going forward.

The goal of this article is to serve as a news-style, in-depth account and analysis of that address. I’ll cover the core announcements, unpack the phrase I used on stage — “graphics three point o” — and explain why digital twins, virtual worlds, and simulation-based approaches are central to the future of robotics and physical AI. I’ll also highlight where breakthroughs are likely to come from, how practitioners can prepare, and the ethical and practical challenges we must confront.

Table of Contents

• 📰 Executive Summary
• 🎤 The Address: Opening Remarks and Core Statements
• 🔬 “So today, we’re gonna talk about simulation.”
• 🤖 “Everything that moves is going to be a robot.”
• 🌐 “AI is transforming graphics, and graphics is transforming AI.”
• 🔮 We believe we are now in Graphics 3.0
• 🧭 Digital twins and virtual worlds as foundational infrastructure
• ⚙️ How simulation powers Physical AI and robotics
• 🧩 Technical breakthroughs I highlighted
• 🖥️ Use cases across industries
• 🧠 Research priorities and open challenges
• 📈 Immediate recommendations for practitioners
• 🌍 Broader societal implications
• 🔭 Looking ahead: what I expect to see at future SIGGRAPHs
• 📚 Case studies and example projects
• 🧾 A roadmap for researchers and leaders
• 🛡️ Safety-first principles
• 🧾 Conclusion: Where do you go from here?
• ❓ FAQ
• 📝 Final remarks

📰 Executive Summary

At SIGGRAPH 2025 I emphasized three themes that will define the next phase of visual computing and embodied intelligence:

• AI and graphics are now mutually transformative: new machine learning methods are changing how we create images and simulations, and advances in graphics are enabling richer training environments and learned models.
• We are entering “Graphics 3.0”: a new era characterized by foundational breakthroughs that integrate physical realism, scalable simulation, and learned models to power both visual experiences and embodied agents.
• Physical AI (physical intelligence implemented by robots, drones, and other agents) depends on digital twins and virtual worlds; as those environments become richer, so too will the capabilities of real-world systems.

Those are the headlines. Below I present a detailed account of what we discussed, why it matters, and what should come next for researchers, industry leaders, and practitioners.

🎤 The Address: Opening Remarks and Core Statements

We began the session with straightforward, declarative remarks meant to set the tone for the conference. I opened by greeting the community: “Hello, SIGGRAPH. It’s great to be with you.” Aaron followed by narrowing our focus: “So today, we’re gonna talk about simulation.” Then Ming-Yu voiced the central thesis that guided the rest of the address: “AI is transforming graphics, and graphics is transforming AI.”

Those lines aren’t just soundbites. Each one signals a larger trend that I believe will dominate research agendas and commercial roadmaps for the foreseeable future. Let me unpack each statement.

🔬 “So today, we’re gonna talk about simulation.”

Aaron’s opening line—“So today, we’re gonna talk about simulation”—was intentionally focused. Simulation is no longer a niche concern reserved for offline rendering, special effects, or niche engineering tools. Instead, simulation is rapidly becoming the lingua franca that connects generative models, physical learning, and human creativity.

Why simulation matters now

• Training environments: Reinforcement learning and imitation learning thrive when agents can experience millions or billions of diverse scenarios. Real-world data collection is slow, costly, and often unsafe; simulation offers scale and control.
• Bridging perception and action: High-fidelity simulators can provide realistic sensor data (RGB, depth, tactile, LIDAR), enabling models that learn end-to-end visuomotor behaviors.
• Differentiable physics: When physics engines become differentiable, we can backpropagate through dynamics, enabling new forms of learning and system identification that were previously impractical.
• Content creation at scale: Artists, designers, and studios can leverage simulation to iterate faster on effects, lighting, and materials in ways that reduce cost and increase creative freedom.

In short, simulation is the connective tissue between graphics and robotics, between synthetic data generation and real-world deployment. The rest of this article explains how that connective tissue will evolve.

🤖 “Everything that moves is going to be a robot.”

One of the bolder lines that resonated with the audience was the assertion, “Everything that moves is going to be a robot.” I attributed that sentiment to our panel discussion, and it encapsulates how I view the trajectory of automation and embodied AI.

Interpreting the statement

“Everything that moves is going to be a robot” is both literal and provocative. Literally, we are already seeing a surge in autonomous or semi-autonomous systems across domains: warehouse robots, delivery drones, surgical assistants, autonomous vehicles, and robotic camera rigs in film production. Provocatively, the statement challenges us to think about the scope of robotics not as isolated platforms, but as a paradigm shift where dynamic systems—whether industrial conveyors, medical devices, or interactive toys—are endowed with sensing, computation, and control.

Consequences for industry and design

• Product design will incorporate autonomy from the outset: mechanical design, software stacks, and safety systems will be co-developed.
• Simulation-driven validation will be mandatory: verifying billions of possible interactions in simulation before release will shorten development cycles and reduce field failures.
• Human-robot interaction becomes central: as more systems move in shared spaces with people, perception, intent prediction, and socially-aware planning become crucial research topics.

I don’t mean to suggest that every moving object will literally have a humanoid form or take complex actions. Instead, the point is that motion will be intelligent. Control loops, adaptive behaviors, and learned policies will permeate moving systems—hence, everything that moves will effectively be a robot.

🌐 “AI is transforming graphics, and graphics is transforming AI.”

Ming-Yu framed what I view as the defining feedback loop of the coming decade: machine learning is reinventing how we generate, manipulate, and think about images and scenes, while state-of-the-art graphics systems give AI the simulated worlds it needs to scale.

How AI transforms graphics

• Generative models are revolutionizing content creation: diffusion models, GANs, autoregressive transformers, and hybrid architectures now create images, textures, and even 3D geometry at unprecedented quality.
• Neural rendering is reshaping pipelines: learned radiance fields (NeRFs) and neural textures are simplifying acquisition and rendering, enabling view synthesis from sparse captures.
• Material and lighting modeling are becoming data-driven: learned BRDFs and illumination models reduce the need for manual artist-tuning and produce more realistic results in varied conditions.

How graphics transforms AI

• Virtual worlds provide abundant training data: high-fidelity, parameterized environments supply the diversity and edge-case exposure that supervised and reinforcement methods demand.
• Physically plausible simulators support reality transfer: as simulators close the realism gap, policies trained in synthetic worlds transfer more reliably to the real world.
• Interactive rendering enables differentiable training: when rendering pipelines are differentiable, learning can incorporate perceptual losses directly in the loop.

Put together, this feedback loop creates a virtuous cycle: better graphics enable better AI, which in turn produces new tools for graphics. This is what I mean by “graphics three point o.”

🔮 We believe we are now in Graphics 3.0

“We believe we are now in graphics three point o,” I stated deliberately during the address. That phrase is not marketing jargon—it’s a classification. If Graphics 1.0 was the era of rasterization and hardware acceleration that democratized 3D, and Graphics 2.0 was the rise of physically based rendering and real-time pipelines that brought cinematic quality to games, then Graphics 3.0 is the phase where AI-driven models, scalable simulation, and continuous learning reshape the rules for what’s possible.

Defining features of Graphics 3.0

1. Integration of learning and physics: Graphics engines will incorporate learned components alongside physics-based modules, producing hybrid systems that are both data-efficient and physically plausible.
2. Simulation at unprecedented scale: Worlds will be procedurally and plausibly generated to provide diverse training scenarios for both visual tasks and embodied agents.
3. Interactive and adaptive pipelines: Content creation and rendering workflows will be AI-augmented, automating routine tasks while allowing creators to intervene with semantic intent.
4. Convergence across modalities: Visual, tactile, and auditory simulations will be tightly coupled, enabling multimodal agents and richer experiences.

Examples illustrating Graphics 3.0

To make Graphics 3.0 concrete, consider a few scenarios:

• A production studio uses learned neural materials and photorealistic simulation to generate background plates that are indistinguishable from real footage. The same simulator is reused to train aerial drones that must navigate in that environment.
• An automaker employs a procedurally generated virtual city to train autonomous driving policies. The simulator includes physically accurate vehicle dynamics, sensor noise models, and unpredictable human behaviors derived from reinforcement learning.
• A research lab builds an interactive avatar creation system: a user provides a few photos and a brief verbal description, and the system constructs a rigged, animatable 3D character with believable cloth dynamics and expressive facial controls—all by combining neural rendering with differentiable simulation.

These are not hypotheticals; they are near-term outcomes as the foundational pieces fall into place.

🧭 Digital twins and virtual worlds as foundational infrastructure

One of the central claims I made is that digital twins and virtual worlds are foundational to physical AI and the future of robotics. I want to expand on that claim and explain why I see them as infrastructure-level technologies.

What I mean by digital twins and virtual worlds

Digital twins are high-fidelity, parameterized models of specific real-world systems—factories, vehicles, buildings, or even individual machines. Virtual worlds are broader: persistent, interactive environments that can represent cities, interiors, nature, or hybrids for training, testing, and content creation.

Why they are foundational

• Reproducibility and safety: Digital twins enable systematic testing of policies and components under controlled conditions without risk to people or infrastructure.
• Continuous monitoring and feedback: A deployed robot can be paired with a digital twin to simulate hypothetical actions, predict failures, and plan maintenance.
• Scalability: Virtual worlds can spawn countless variations of scenarios to stress-test models for edge cases that would be infeasible or dangerous to reproduce in reality.
• Cross-domain transfer: Shared virtual worlds allow teams across design, engineering, and operations to work with a common representation—reducing friction between disciplines.

Digital twins are not merely visualization tools; they are living models that interact with deployed systems in closed-loop ways. When an industrial line changes a component, the digital twin updates and simulates the impact on throughput and safety. When a robot in the field encounters a rare environment, an accurate virtual twin can be used to replay and analyze the event, refine policies, and deploy improved controls.

⚙️ How simulation powers Physical AI and robotics

Simulation is central to what I called “physical AI”—AI that produces intelligent behavior in the physical world. Below I outline the major simulation capabilities that will matter most for robotics and embodied systems.

High-fidelity sensor simulation

Robotic systems rely on sensors: cameras, depth sensors, LIDAR, IMUs, tactile arrays. For agents to learn robust policies, simulators must model sensor characteristics faithfully, including noise characteristics, latency, lens distortions, motion blur, and occlusion. Advances in neural rendering and physics-based modeling make it feasible to create sensor outputs that closely match real-world measurements.

Accurate dynamics and contact modeling

Contact-rich interactions—grasping, manipulation, locomotion—are traditionally difficult to model. Differentiable physics engines, hybrid analytic-learned contact models, and data-driven friction models are converging to produce simulators that can replicate subtle interactions between objects and agents. This fidelity is crucial for transferring manipulation policies to real robots.

Environment variability and procedural generation

A policy trained on a single environment will overfit. Procedural generation introduces variability at scale—randomized lighting, textures, object placement, and behaviors—which drives model robustness. Crucially, procedural generation must be plausible: domain randomization is effective only when variability remains within the distributional envelope of real-world conditions.

Multimodal and social simulation

Robots will operate alongside people. Simulators must therefore account for social norms, predictive human behavior, and multimodal interactions (speech, gesture, gaze). Synthetic people, animated by AI models trained on large-scale behavioral data, can provide agents with realistic cohabitation scenarios.

Scale and compute efficiency

Training modern models requires huge compute budgets. Efficient simulation—through parallelization, model reduction, and learned approximations—will be essential to close the loop between graphics and AI in a cost-effective way.

🧩 Technical breakthroughs I highlighted

During the address I sketched a number of technical breakthroughs that we, and other labs, are pursuing to make Graphics 3.0 a reality. Below I outline these areas and explain why each is pivotal.

1. Differentiable simulation and rendering

Differentiability changes everything. When physics engines and renderers expose gradients, they enable end-to-end optimization of systems that span perception, planning, and control. For example, a robot can refine its model of object mass or surface friction by comparing predicted and observed trajectories and backpropagating errors through a differentiable simulator.

2. Neural scene representations

Representations like Neural Radiance Fields (NeRF) and subsequent variants let us model complex geometry, appearance, and view-dependent effects in compact, learned forms. They are enabling new capture technologies and interactive rendering modes. As these representations become more dynamic and physically grounded, they will be used not just for rendering but for physical simulation and planning as well.

3. Generative models for environment and agent behavior

Generative models—diffusion models, transformers, autoregressive architectures—are now powerful enough to synthesize textures, shapes, and even behavior. We can use these models to populate virtual worlds with realistic objects and agents whose behavior is learned from large datasets, creating more realistic training scenarios.

4. Self-supervised and representation learning for control

Self-supervised learning reduces the need for labeled datasets by learning structure from raw interactions. For robotics, this means agents can bootstrap robust features for perception and control from passive data collection, which is often available at scale.

5. Sim-to-real transfer and domain adaptation

Better sim-to-real methods—domain randomization, adversarial adaptation, and learned correction policies—are closing the gap between virtual training and real-world deployment. Combined with online adaptation mechanisms, agents can continue to refine policies after deployment with minimal supervision.

6. Scalable multi-agent and social simulations

Large-scale, multi-agent environments enable the study of emergent behaviors, collaboration, and competition. These simulations are pertinent not only for robotics but for understanding complex systems ranging from traffic to logistics.

🖥️ Use cases across industries

Graphics 3.0 and simulation-driven physical AI will affect many industries. Here I highlight a set of domains where I expect immediate, high-impact applications.

Entertainment and media

• Real-time volumetric capture and neural rendering will change how films and live events are produced.
• Procedural virtual worlds populated by AI agents can serve as dynamic backdrops for games and interactive experiences.
• Automated, AI-assisted content creation reduces iteration time for studios and indie creators alike.

Automotive and mobility

• Virtual cities and sensor-accurate driving simulators allow safe training and validation of autonomy systems across rare edge cases.
• Digital twins of roads and infrastructure support fleet optimization and predictive maintenance.

Manufacturing and logistics

• Robots trained in simulated warehouses can perform complex pick-and-place tasks with higher generalization.
• Digital twins of factories help optimize throughput, reduce downtime, and simulate human-robot collaboration for safety planning.

Healthcare and medical devices

• Surgical simulations with realistic tissue models can train both algorithms and clinicians.
• Robotic assistants tested in virtual operating rooms reduce risk during early deployments.

Defense and aerospace

• High-fidelity flight and mission simulators driven by learned models improve pilot training and autonomous mission planning.
• Virtual testing grounds allow safe stress-testing of autonomy under extreme conditions.

🧠 Research priorities and open challenges

To make the promise of Graphics 3.0 real, the research community must solve several hard problems. I outlined these as priorities during the presentation.

Bridging the reality gap

Despite dramatic improvements, gaps remain between simulation and reality. We need principled methods to quantify and shrink that difference, either by improving simulators or by making policies robust to the remaining mismatch.

Data efficiency and generalization

Many models achieve impressive results but require massive data and compute. Developing approaches that generalize from limited interactions is critical for resource-constrained deployments and for settings where safety is paramount.

Long-horizon planning and hierarchical learning

Tasks that require extended temporal reasoning—assembly, inspection, or multi-stage manipulation—demand hierarchical architectures that combine high-level planning with low-level control. Integrating those levels in simulation and training pipelines remains a core challenge.

Ethics, safety, and governance

As virtual worlds become more realistic and robots more autonomous, ethical concerns become central. We must develop standards for safety, privacy, and accountability, and ensure that simulated data does not unintentionally encode biases or misrepresent vulnerable communities.

Compute and energy efficiency

Training at the scale necessary for these systems consumes large amounts of compute and energy. Research into more efficient architectures, sparse computation, and hardware-software co-design will be vital.

📈 Immediate recommendations for practitioners

If you are building systems that will benefit from Graphics 3.0, here are practical steps I recommended at SIGGRAPH and expand on here.

Adopt simulation early in the development cycle

Use virtual prototypes to explore design tradeoffs and safety cases before building hardware. Early simulation reduces costly physical iterations and uncovers failure modes.

Design for data diversity

When creating synthetic datasets, prioritize diversity in lighting, geometry, texture, and agent behavior. Too little variability is often the single biggest cause of poor transfer.

Invest in differentiable components where possible

Make key parts of your pipeline differentiable—sensors, dynamics, and rendering—so you can perform gradient-based calibration and system identification.

Co-design hardware and software

If your system requires specialized sensors or actuators, consider how hardware choices influence simulation fidelity and the difficulty of sim-to-real transfer.

Establish clear safety validation metrics

Define tests in simulation that correspond to safety and reliability goals in the field. Track those metrics continuously as you iterate models and environments.

🌍 Broader societal implications

As systems trained in virtual worlds become increasingly integrated into daily life, we must confront cultural, economic, and policy implications.

Workforce transformation

Automation will change job roles across industries. While many tasks will be augmented by robots, humans will continue to be essential in supervision, creative direction, and exception handling. Education programs must evolve to teach simulation literacy, AI ethics, and system integration skills.

Equity and accessibility

High-fidelity simulation platforms and compute resources are currently concentrated among large organizations. Democratizing access to these tools—through open-source initiatives, shared datasets, and cloud credits—will ensure broader participation and innovation.

Environmental impact

Training at scale consumes energy. We must invest in energy-efficient algorithms, improved hardware, and renewable-powered datacenters to mitigate environmental costs.

Regulation and public trust

Governments and standards bodies need to engage with researchers and industry to create frameworks for safety testing, certification, and public oversight. Transparent benchmarks and reproducible validation pipelines will build trust in autonomous systems.

🔭 Looking ahead: what I expect to see at future SIGGRAPHs

I closed the address by predicting a surge of new ideas at SIGGRAPH in the coming years—and I stand by that forecast. Here is what I expect to be on display at SIGGRAPH 2026 and beyond.

• End-to-end differentiable content creation tools that let artists optimize scenes by specifying high-level objectives.
• Open, simulated cities used as shared benchmarks for autonomous mobility and urban robotics.
• Hybrid renderer-simulator stacks that produce photorealistic sensor outputs while supporting physical interactions for robotics research.
• Large-scale, multimodal datasets collected in virtual worlds to train agents for tasks that require joint visual, tactile, and audio reasoning.
• New standards for safety evaluation and sim-to-real certification that enable regulatory compliance for autonomous systems.

These developments will not happen in isolation. They will be driven by a combination of academic research, industrial investment, and open-source collaboration.

📚 Case studies and example projects

To ground the discussion, I’ll describe a few hypothetical but realistic projects that typify the Graphics 3.0 approach. Each example integrates simulation, learned components, and physical deployment.

Case study 1: Warehouse automation with digital twins

A logistics company builds a digital twin of its fulfillment center. The twin models conveyor dynamics, robot kinematics, and worker movement. Policies for mobile picking robots are trained in the twin under thousands of procedural variations: different pallet densities, lighting conditions, and rare failure modes like dropped items.

After extensive simulated validation, the robots are deployed with an online adaptation layer that uses real sensor feedback to fine-tune grasping policies. The result is a 40% reduction in order fulfillment errors and a 25% improvement in throughput.

Case study 2: Film production using neural rendering

A studio shoots live action on a soundstage augmented by a neural backdrop generated from sparse captures of exotic locations. Neural materials and learned lighting ensure plausible integration of actors and CG elements. Because much of the environment is rendered rather than physically built, the studio iterates lighting and camera moves in real time, reducing production time and costs.

Case study 3: Autonomous drone navigation with multimodal simulation

A robotics startup develops autonomous drones for search-and-rescue. They train perception and planning modules in a virtual mountain environment that includes realistic wind, fog, and occlusion. The simulator models the drone’s IMU and camera pipeline, and a learned wind model introduces realistic disturbances. The team transfers policies to hardware using a small domain adaptation dataset and achieves robust performance in early field tests.

🧾 A roadmap for researchers and leaders

If you lead a lab, a product team, or a startup, here is a practical roadmap you can follow to align with Graphics 3.0.

Short-term (0-12 months)

• Adopt modular simulation stacks and begin producing synthetic datasets for core tasks.
• Benchmark sim-to-real gaps on small-scale problems and document failure modes.
• Invest in differentiable components for system identification and calibration.

Medium-term (1-3 years)

• Develop procedural world generators and integrate learned generative models to create diverse environments.
• Build pipelines for continuous integration of simulation tests with real-world deployment monitoring.
• Collaborate with standards bodies and the research community to define safety benchmarks.

Long-term (3+ years)

• Contribute to and adopt common virtual world standards that enable composability of environments and models.
• Scale training to multi-agent, multi-modal systems and prioritize energy-efficient architectures.
• Support workforce reskilling programs and open resource initiatives to democratize access to simulation tools.

🛡️ Safety-first principles

Throughout our work we must keep safety at the center. Here are guiding principles I advocate:

• Fail-safe by design: design systems so that failure modes are graceful and predictable.
• Transparent validation: publish benchmarks, test suites, and evaluation code whenever possible.
• Human-in-the-loop operation: maintain mechanisms for human oversight and intervention, especially in high-risk settings.
• Continuous monitoring: use digital twins and telemetry to detect degradation and trigger retraining or rollback.

Safety is not a single checkpoint; it’s an ongoing engineering commitment throughout a system’s lifecycle.

🧾 Conclusion: Where do you go from here?

What comes next? For those attending SIGGRAPH and those reading this report, my advice is pragmatic: invest in simulation capabilities, embrace the convergence of AI and graphics, and think holistically about the integration of perception, control, and rendering. I expect to see an explosion of new ideas at SIGGRAPH in the coming years because the field is wide open for fundamental breakthroughs again.

Digital twins and virtual worlds are not peripheral—they are foundational to the next generation of robotics and physical AI. If you are a researcher, explore differentiable physics and neural scene representations. If you are a practitioner, prioritize simulation-first development and safety validation. If you are a leader, commit resources to the tooling and compute infrastructure that makes Graphics 3.0 possible.

I’m excited by the opportunities and the challenges ahead. I believe Graphics 3.0 will reshape industries and research in profound ways, and SIGGRAPH will be a key venue where the most consequential ideas are shared, critiqued, and advanced.

❓ FAQ

Q1: What exactly do you mean by “Graphics 3.0”?

Graphics 3.0 is the next phase in the evolution of visual computing. It’s characterized by the deep integration of learned models, differentiable simulation, and physically plausible rendering. In practice that means hybrid systems where neural networks and physics engines work together to produce realistic visuals and to enable agents to learn in simulated environments that closely mimic reality.

Q2: Are digital twins and virtual worlds the same thing?

They are related but not identical. A digital twin is typically a high-fidelity, parameterized model of a specific real-world system used for monitoring, simulation, and planning. A virtual world is a broader, often persistent environment used for training, testing, or entertainment and may represent cities, natural environments, or fictional terrains. Both are essential, but digital twins are often more closely coupled with deployed systems.

Q3: How do you ensure sim-to-real transfer works?

There are multiple strategies: improve simulator fidelity, use domain randomization to expose policies to a wide variety of scenarios, apply adversarial or feature-level domain adaptation methods, and incorporate online adaptation post-deployment. Differentiable simulation and data-driven correction models also help reduce mismatch.

Q4: Won’t this increase centralization of power among large companies with compute?

There is a risk of centralization because high-fidelity simulators and large-scale training can require significant resources. To counter this, I advocate for open-source tools, shared benchmarks, and cloud credits that democratize access. Industry-academic partnerships and community-driven datasets can also help level the playing field.

Q5: What are the main ethical concerns?

Key concerns include safety, privacy, surveillance misuse, bias in simulated or learned behaviors, and the socio-economic impacts of automation. Addressing these issues requires multi-stakeholder engagement—researchers, policymakers, civil society, and industry must collaborate to create standards and governance frameworks.

Q6: How should small teams start experimenting with these ideas?

Start with modular, open-source simulation tools and focus on low-dimensional problems where you can validate sim-to-real transfer. Use procedural generation to create varied datasets, and test differentiable components on calibration tasks. Leverage community resources and collaborate with academic partners where possible.

Q7: What role will SIGGRAPH play in this evolution?

SIGGRAPH is a crucible for cross-disciplinary innovation in graphics, visualization, and interactive techniques. As simulation and AI become central to visual computing, SIGGRAPH will become an even more important venue for sharing breakthroughs that span rendering, perception, and robotics. Expect to see more papers and demos that bridge these areas.

Q8: How long until Graphics 3.0 impacts everyday products?

Some aspects are already impacting products—neural rendering in rendering pipelines, simulation-driven robotics in warehouses, and generative content tools for media. Wider impacts, such as pervasive digital twins and large-scale virtual cities for autonomy, will likely mature over the next 3–7 years as tools, standards, and compute become more accessible.

Q9: What skills should students and new researchers cultivate?

Learn the fundamentals of graphics (rendering, shading, materials), machine learning (especially generative models and reinforcement learning), and physics/dynamics. Familiarity with differentiable programming, simulation frameworks, and systems engineering will be invaluable. Interdisciplinary fluency—being able to speak to both graphics and robotics experts—will be a major asset.

Q10: How can the community contribute to safer development?

Share reproducible benchmarks, publish negative results, build open validation suites, and engage with policymakers to develop standards. Contribute to public datasets and toolkits that enable smaller teams to test safety-critical scenarios without prohibitive expense.

📝 Final remarks

At SIGGRAPH I said that AI is transforming graphics and graphics is transforming AI. That statement is more than a slogan: it’s a roadmap. Over the next several years I expect to see intense cross-pollination between these fields, with simulation and digital twins at the center. I’m optimistic about the innovations to come and clear-eyed about the responsibilities we bear as creators and deployers of these technologies.

If you’re building the next generation of tools, models, or robotic systems, my advice is straightforward: invest in simulation, prioritize safety and reproducibility, and embrace the interdisciplinary collaboration that Graphics 3.0 demands. SIGGRAPH will be one of the stages where this transformation unfolds, and I look forward to working with the community to turn these ideas into robust, beneficial technologies.

---