NVIDIA's ICRA 2026 Research: How Simulation-Trained Robots Are Mastering the Real World

Robotics is hitting an inflection point. For years, impressive demos have been confined to labs and tightly scripted factory floors. Now, a wave of new research from NVIDIA suggests that simulation-trained robots are beginning to function reliably in messy, unpredictable real-world settings. At the 2026 International Conference on Robotics and Automation (ICRA), NVIDIA Research presented 28 accepted papers—eight of which specifically demonstrated how simulation-to-real (sim-to-real) transfer is helping robots perceive, reason, plan, and act in dynamic environments .

The throughline is unmistakable: training in high-fidelity simulation, rather than painstakingly collecting millions of real-world demonstrations, is becoming the scalable foundation for generalizable, reliable embodied autonomy outside the laboratory .

The Eight Papers: A Full-Stack Push into the Real World

The eight papers collectively address the core challenges robot developers face today, from multi-arm coordination to vision-language-action reasoning.

ScheduleStream: 3x Faster Multi-Arm Planning

Traditional robot scheduling software processes arms sequentially, creating bottlenecks in multi-arm cells. ScheduleStream runs computations on GPUs, enabling multiple arms to plan movements and operate in parallel. Running on the NVIDIA Jetson edge AI platform, it delivered a 3x speedup across multi-arm planning scenarios. The framework is open source on GitHub .

COMPASS: Cross-Embodiment Navigation with 80% Real-World Success

Building robots that navigate across different body types—wheeled mobile robots, humanoids—is notoriously hard. The COMPASS policy framework first trains a baseline navigation policy via imitation learning, then uses residual reinforcement learning in NVIDIA Isaac Lab to create specialized policies for diverse embodiments—all in simulation. Compared with imitation learning baselines, COMPASS achieved a 4.5x improvement in average success rate. It also transferred seamlessly to the real world, demonstrating ~80% success across 20 real-world navigation trials on autonomous mobile robots and humanoids .

Grasp-MPC: Adaptive Grasping That Corrects Mid-Motion

Fixed grasping plans fail when objects shift or when a robot's initial estimate is slightly off. Grasp-MPC continuously corrects the robot's motion as it closes in on an object. Researchers generated 2 million simulated trajectories across 8,000 objects using the GraspGen dataset and cuRobo, a CUDA-accelerated motion generation library. On real robots, it achieved ~75% overall grasping success versus a 41% baseline .

Deformable Cluster Manipulation: Zero-Shot Clutter Clearing

Manipulating tangled, flexible materials—like branches over power lines—requires more than a precision gripper. NVIDIA researchers trained policies to use the entire arm to sweep aside clusters, using thousands of synthetic trees in Isaac simulation frameworks. The result: policies deployed zero-shot to real branches with no additional training .

PEEK: Focused Vision Delivers a 41x Accuracy Boost

Distractors in a robot's camera feed can derail even well-trained manipulation policies. PEEK uses a vision-language model to read a task instruction and focus the robot's vision on relevant objects while fading out everything else. When added to a policy trained purely in simulation, PEEK produced a 41x real-world improvement in accuracy. For large vision-language-action (VLA) models, gains ranged from 2–3.5x. PEEK integrates with any camera-based policy without modification .

Do What You Say (SEAL): Closing the Reasoning-Action Gap

The SEAL framework—a collaboration with Carnegie Mellon University, the University of Utah, and the University of Sydney—fixes a deceptively common failure mode: the model reasons correctly, picks the right plan, but executes something different. SEAL generates multiple candidate action sequences, simulates where each leads, and selects the one that best matches the stated intent. It delivers up to 15% accuracy gains over prior work and is robust to rephrased instructions, clutter, and changed camera angles .

Refinery: Chaining Assembly Steps with 91% Simulation Success

For multi-part assemblies, each step's outcome shapes the next. Refinery trains policies that understand these dependencies, learning across hundreds of simulated scenarios. It achieves 91% simulation success and a nearly 11% mean improvement over baselines, with policies that chain together for long, complex assembly sequences .

Vision-Based Sim-to-Real RL for Dexterous Humanoid Tasks

A separate vision-based sim-to-real reinforcement learning recipe trained a humanoid robot for grasp-and-reach, box lift, and bimanual handover tasks. The approach demonstrated high success rates on unseen objects with robust, adaptive behaviors—underscoring that vision-based dexterous manipulation via sim-to-real RL is both viable and scalable .

The Underlying Platforms: GR00T, Cosmos, Newton, cuMotion, and Jetson

The eight papers rest on several cross-cutting NVIDIA platforms that turn simulation into a practical, end-to-end development environment:

• NVIDIA Isaac GR00T: A research initiative and development platform for general-purpose robot foundation models, featuring open models (the GR00T N series) for humanoid reasoning and control, simulation frameworks built on Omniverse and Cosmos, and the Jetson AGX Thor edge computer .
• NVIDIA Cosmos World Models: A platform of state-of-the-art generative world foundation models for building custom world models at scale, supporting synthetic data generation and policy evaluation entirely in simulation .
• Newton 1.0 Physics Engine: An open-source, GPU-accelerated physics engine co-developed with Google DeepMind and Disney Research, now available in NVIDIA Isaac Lab to help create more capable and adaptable robots .
• NVIDIA cuMotion: A CUDA-accelerated library that solves robot motion planning problems at scale by running multiple trajectory optimizations simultaneously to return the best solution .
• NVIDIA Jetson Edge AI Platform: The hardware platform used aboard robots to run the full robot stack, including the ScheduleStream 3x speedup demo, with Jetson AGX Thor serving as the onboard computer .

Partners Putting Sim-to-Real Efficiency to Work

Toyota Research Institute Cuts Real-World Data Needs

Toyota Research Institute (TRI) customized NVIDIA Cosmos world foundation models for dynamic view synthesis and robot teleoperation, reducing the amount of real-world data needed to train vision-based manipulation policies .

Mimic Robotics: 10x Better Sample Efficiency

Mimic Robotics developed a video-action model using NVIDIA's platforms that achieves 10x better sample efficiency and 2x faster convergence on real-world manipulation tasks, dramatically cutting the number of expensive real-world demonstrations required .

Doosan Robotics Adds Context-Aware Palletizing

Doosan uses NVIDIA Cosmos Reason to let palletizing robots analyze the contents of boxes, detect damage, and adjust handling based on weight and fragility—enabling context-aware decision-making without exhaustive real-world training data .

The Bigger Picture: From Scripted Demos to Embodied Autonomy

NVIDIA framed this body of work as part of a fundamental shift in the robotics industry:

"Robotics is entering a new phase: moving from controlled demos and scripted automation toward generalizable, reliable embodied autonomy in the real world" .

Sim-to-real transfer is no longer an academic curiosity. The eight ICRA papers show it tackling the full stack: parallel multi-arm coordination, cross-embodiment policy generalization, novel-object grasping in clutter, zero-shot deformable manipulation, precise sequential assembly, and vision-language-action models that reason before they move . The clear message: simulation-based training—rather than reliance on vast quantities of real-world human demonstrations—is the scalable path to robots that function robustly in unstructured, dynamic environments.