Top 10 AI models
The gap between the capabilities of language models and automated publishing has narrowed significantly over the past 18 months. A new class of core models-designed specifically not for text generation but for physical work-now runs on real hardware across factories, warehouses, and research labs. These systems include published robot policies, private preview VLAs, open weight search models, and global models used to scale robot training data. Some are being evaluated or deployed with industrial partners; Others are primarily research or developer-oriented systems. Here’s a breakdown of the ten most important in 2026.
Nvidia Isaac GR00T N-Series (N1.5/N1.6/N1.7)
NVIDIA released the original GR00T N1 at GTC in March 2025 as the world’s first open, fully customizable core model for generalized human thinking and skills. Since then the N series has progressed rapidly. GR00T N1.5, announced at COMPUTEX in May 2025, provided key improvements to VLM and Eagle 2.5, a FLARE training target that enables learning from human ego videos, and the GR00T-Dreams scheme – which reduced synthetic data generation from months to approximately 36 hours.
This was followed by the GR00T N1.6 on December 15, 2025, with a new internal NVIDIA Cosmos-2B VLM backbone supporting flexible resolution, 2x larger DiT (32 layers versus 16 in the N1.5), case-relative motion parts for smoother movement, and several thousand additional hours of teleoperation data from YAM two-handed joysticks, AGIBot Genie-1, and Unitree G1. It has been validated in real manual and motor processing tasks via these paradigms.
The latest release, GR00T N1.7 Early Access (April 17, 2026), is an open, commercially licensed 3B-parameterized VLA built on a Cosmos-Reason2-2B backbone with 32 DiT layers for low-level drive control – a dual-system Action Cascade architecture. Its central offering is EgoScale: pre-training on 20,854 hours of egocentric human video covering more than 20 task categories, significantly exceeding the hours of robot teleoperation used in previous releases. NVIDIA has set what it describes as the first-ever law for measuring robot prowess – 1,000 to 20,000 hours of selfish human data more than double the average task completion. N1.7 Early Access is available on HuggingFace and GitHub under an Apache 2.0 license, with full production support tied to the general availability release. Early adopters across the GR00T N series include AeiRobot, Foxlink, NEURA Robotics, and Lightwheel.
Google DeepMind Gemini Robotics 1.5
Gemini Robotics is an advanced vision, language and movement (VLA) model built on Gemini 2.0, adding physical actions as a new output method for direct control of robots. It was launched in March 2025 alongside Gemini Robotics-ER (Embodied Reasoning). The September 2025 update, Gemini Robotics 1.5, introduced agentic capabilities – converting visual information and instructions into motor commands while making the model’s thought process transparent, helping robots evaluate and complete complex multi-step tasks more clearly.
Access remains available to select partners including Agile Robots, Agility Robotics, Boston Dynamics, and Enchanted Tools, and is not available to the general public. The broader family continues to evolve: Gemini Robotics-ER 1.6, released on April 14, 2026, enhances spatial reasoning and understanding multiple perspectives – including the ability to read new instruments developed in collaboration with Boston Dynamics to read complex measuring devices and vision glasses. Gemini Robotics-ER 1.6 is available to developers via the Gemini API and Google AI Studio.
Physical intelligence π0 / π0.5 / π0.7
π0 proposes a flow matching architecture built on top of a pre-trained vision language model for Internet-scale semantic knowledge inheritance, which has been trained across multiple robotic platforms including single-armed robots, two-armed robots, and mobile manipulators. Open source physical intelligence π0 in February 2025.
π0.5 was published on April 22, 2025, with openpi weights released later in 2025. Rather than targeting improved dexterity, it focuses on open-world generalization: the model uses cross-training across heterogeneous tasks, multiple robots, high-level semantic prediction, and web data to clean unfamiliar kitchens and bedrooms that were not seen in training. A later version applied the RECAP approach (RL with Experience and Corrections by Advantage-Contingent Policies) – training by demonstration, training by corrections, and improvement by independent experience – for which physical intelligence reported doubling productivity on tasks such as inserting a filter into an espresso machine, folding never-before-seen laundry, and assembling a cardboard box.
The most recent public research release is π0.7, published on April 16, 2026. It is a research phase system that focuses on synthetic generalization: combining skills learned from different contexts to solve tasks for which the model has not been explicitly trained. He describes physical intelligence as a steerable model with emergent capabilities, an early but meaningful step toward a general-purpose robotic brain. The paper uses subtle hedging language throughout, and no timeline for commercial publication is mentioned.
Figure AI Helix
Released on February 20, 2025, the Helix is the first VLA device to produce continuous, high-speed control of the entire human upper body, including the wrists, torso, head and individual fingers. A dual system design is used: System 2 is an online pre-trained VLM with parameter 7B and operates at 7-9 Hz for scene understanding and language understanding; System 1 is an 80M cross-parameter encoder and decoder operating at 200 Hz, translating S2 semantic representations into continuous and precise robotic actions. The model was trained on approximately 500 hours of data operated remotely by multiple robots and multiple operators, with automatic tagging of instructions via VLM applied too late. All training components are excluded from assessments to prevent contamination.
Helix powers fully integrated low-power integrated GPUs, making it relevant for commercial publishing research and future applications for humans. It uses a single set of neural network weights for all behaviors-picking and placing items, using drawers and refrigerators, and interaction between robots-without any fine-tuning for a specific task. It has been proven in household processing tasks and logistical package sorting, and can operate simultaneously on two robots through a supervisory architecture that divides overall objectives into subtasks for each robot.
OpenVLA
OpenVLA is an open source 7B-parameter VLA trained on a variety of 970,000 real-world robot demonstrations. It is based on the Llama 2 language model combined with a visual encoder that incorporates pre-trained features from DINOv2 and SigLIP. Despite being 7 times smaller, OpenVLA outperforms the enclosed RT-2-X (55B parameters) by 16.5 percentage points in absolute mission success rate across 29 missions and multiple robotic incarnations.
A February 2025 paper presented an OFT (optimal fine-tuning) recipe, which combines parallel decoding, action segmentation, continuous action representation, and an L1 regression target. OFT provides 25 to 50 times faster inference speed and achieves an average success rate of 97.1% on the LIBERO simulation benchmark, outperforming π0, Octo, and the propagation policy. The enhanced version, OFT+, adds FiLM adaptation to improve language foundations and enables high-frequency manual control of the ALOHA robot. OpenVLA supports fine-tuning and quantization of LoRA for resource-constrained deployment, and there are community-based ROS 2 wrappers for integration with robot operating systems.
Octo
Octo is an open source public robot policy from UC Berkeley, available in two sizes: Octo-Small (27 million parameters) and Octo-Base (93 million parameters). Both use a transformer backbone with diffusion decoder, which is pre-trained on 800,000 bot loops from the Open X-Embodiment dataset across 25 datasets. The model supports both natural language instructions and target image adaptation, and accommodates flexible monitoring and working spaces including new sensors and motion representations without architectural changes.
Octo is specifically designed to support efficient tuning of new robot settings. In formal evaluation, each task uses approximately 100 demonstrations of the target domain, and Octo outperforms training from scratch by an average of 52% across six evaluation settings spanning institutions including CMU, Stanford, and UC Berkeley. It performs similarly to the RT-2-X (55B parameters) at zero shot settings while being smaller. Octo is primarily a research and development tool, and is a powerful, lightweight starting point for labs that need to quickly iterate on new processing tasks with limited compute.
AGIBOT BFM and GCFM
In April 2026, Shanghai-based AGIBOT announced two core models as part of its complete “One Robotic Body, Three Intelligences” architecture. The behavioral foundation model (BFM) is centered around imitation and transfer of behavior, and is designed to efficiently acquire new motor behaviors from demonstrations. The Generative Control Foundation Model (GCFM) is based around generating context-aware robot movements from multimodal inputs including text, audio, and video.
AGIBOT is positioning AGIBOT WORLD 2026 as part of the data foundation for its broader robotics portfolio – a real, open source, production-level dataset covering commercial spaces, homes and everyday scenarios. The company announced 2026 as the “first deployment year” at its partner conference in April 2026 and announced the rollout of its 10,000th robot in March 2026.
Gemini Robotics on the device
Gemini Robotics On-Device is a VLA model for two-arm robotics designed to operate locally on the robot itself with low-latency inference, without the need for a data network connection. Released in June 2025, it is the first VLA model that Google DeepMind has made available for fine-tuning. It builds on the task generalization and dexterity capabilities of the Gemini Robotics cloud-based model, and is optimized for on-device implementation where latency or connectivity constraints apply. The model was primarily trained on ALOHA robots and modified for Apptronik’s Franka FR3 dual-arm and Apollo humanoid robots. It adapts to new tasks with at least 50 to 100 demos. Currently availability is through specific trusted testers, not a generic release.
NVIDIA Cosmos World Foundation Models
Cosmos is not a robotics policy model in the traditional sense, but rather a global generative model that produces synthetic trajectory data to scale the training pipelines of other models on this list. The GR00T-Dreams scheme uses Cosmos software to generate massive amounts of synthetic trajectory data from a single image and a single language instruction, enabling robots to learn new tasks in unfamiliar environments without requiring specific data for remote operation. This directly promoted the development of GR00T N1.5. Cosmos Predict 2, the version used in GR00T-Dreams, is available on HuggingFace with performance improvements for high-quality global generation and reduced hallucinations. Companies including Skild AI and FieldAI use Cosmos and Isaac simulation components to generate training data for synthetic robots and validate robot behaviors in simulation before deploying them in the real world.
SmolVLA (HuggingFace LeRobot)
Released on June 3, 2025, SmolVLA is a 450 million-parameter built-in VLA from HuggingFace built within the LeRobot framework and trained entirely on community-contributed open source data. The vision language backbone uses SmolVLM-2 combined with a flow-matching transformer action expert – to output continuous actions instead of tokens, which is the same action representation used by π0 and GR00T N1. It was pre-trained on 10 million frames curated from 487 community datasets tagged as “lerobot” on HuggingFace, covering diverse environments from laboratories to living rooms.
SmolVLA runs on consumer hardware including RTX-class single GPUs and MacBooks. Official fine-tuning benchmarks show approximately 4 hours on a single A100 for 20,000 training steps. In real robot evaluations using SO100 and SO101 arms, it achieves an average success rate of approximately 78.3% after fine-tuning to a specific task. It matches or outperforms larger models such as ACT on LIBERO and Meta-World simulation benchmarks, and supports asynchronous inference for 30% faster response time and twice the mission throughput. SmolVLA is the most accessible entry point into the VLA ecosystem for teams with limited compute.
