Top 10 Physical AI Models
The gap between language model capabilities and robotic deployment has been narrowing considerably over the past 18 months. A new class of foundation models — purpose-built not for text generation but for physical action — is now running on real hardware across factories, warehouses, and research labs. These systems span deployed robot policies, private-preview VLAs, open-weight research models, and world models used to scale robot training data. Some are being evaluated or deployed with industrial partners; others are primarily research or developer-facing systems. Here is a breakdown of the ten that matter most 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 foundation model for generalized humanoid reasoning and skills. The N-series has since advanced rapidly. GR00T N1.5, announced at COMPUTEX in May 2025, introduced a frozen VLM, Eagle 2.5 grounding improvements, a FLARE training objective enabling learning from human ego videos, and the GR00T-Dreams blueprint — which reduced synthetic data generation from months to approximately 36 hours.
GR00T N1.6 followed on December 15, 2025, with a new internal NVIDIA Cosmos-2B VLM backbone supporting flexible resolution, a 2× larger DiT (32 layers versus 16 in N1.5), state-relative action chunks for smoother motion, and several thousand additional hours of teleoperation data from bimanual YAM arms, AGIBot Genie-1, and Unitree G1. It was validated on real bimanual and locomanipulation tasks across those embodiments.
The most recent release, GR00T N1.7 Early Access (April 17, 2026), is a 3B-parameter open, commercially licensed VLA built on a Cosmos-Reason2-2B backbone with a 32-layer DiT for low-level motor control — an Action Cascade dual-system architecture. Its central advance is EgoScale: pretraining on 20,854 hours of human egocentric video spanning 20+ task categories, significantly scaling beyond the robot teleoperation hours used in prior versions. NVIDIA identified what it describes as the first-ever scaling law for robot dexterity — going from 1,000 to 20,000 hours of human egocentric data more than doubles average task completion. N1.7 Early Access is available on HuggingFace and GitHub with Apache 2.0 licensing, 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-action (VLA) model built on Gemini 2.0, with physical actions added as a new output modality for directly controlling robots. It launched in March 2025 alongside Gemini Robotics-ER (Embodied Reasoning). The September 2025 update, Gemini Robotics 1.5, introduced agentic capabilities — turning visual information and instructions into motor commands while making the model’s reasoning process transparent, helping robots assess and complete complex multi-step tasks more legibly.
Access remains available to selected partners including Agile Robots, Agility Robotics, Boston Dynamics, and Enchanted Tools, and is not publicly available. The broader family continues to evolve: Gemini Robotics-ER 1.6, released April 14, 2026, enhances spatial reasoning and multi-view understanding — including a new instrument reading capability developed in collaboration with Boston Dynamics for reading complex gauges and sight 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 to inherit Internet-scale semantic knowledge, trained across multiple dexterous robot platforms including single-arm robots, dual-arm robots, and mobile manipulators. Physical Intelligence open-sourced π0 in February 2025.
π0.5 was published on April 22, 2025, with openpi weights released later in 2025. Rather than targeting improved dexterity, its focus is open-world generalization: the model uses co-training across heterogeneous tasks, multiple robots, high-level semantic prediction, and web data to clean unfamiliar kitchens and bedrooms not seen in training. A subsequent version applied the RECAP (RL with Experience & Corrections via Advantage-conditioned Policies) approach — training by demonstration, coaching through corrections, and improving from autonomous experience — which Physical Intelligence reported doubled throughput on tasks such as inserting a filter into an espresso machine, folding previously unseen laundry, and assembling a cardboard box.
The most recent public research release is π0.7, published April 16, 2026. It is a research-stage system focused on compositional generalization: combining learned skills from different contexts to solve tasks the model was never explicitly trained on. Physical Intelligence describes it as a steerable model with emergent capabilities — an early but meaningful step toward a general-purpose robot brain. The paper uses careful hedging language throughout, and no commercial deployment timeline has been stated.
Figure AI Helix
Released February 20, 2025, Helix is the first VLA to output high-rate, continuous control of the entire humanoid upper body, including wrists, torso, head, and individual fingers. It uses a dual-system design: System 2 is a 7B-parameter internet-pretrained VLM operating at 7–9 Hz for scene understanding and language comprehension; System 1 is an 80M-parameter cross-attention encoder-decoder transformer running at 200 Hz, translating S2’s semantic representations into precise continuous robot actions. The model was trained on approximately 500 hours of multi-robot, multi-operator teleoperated data, with automatic instruction labeling via a VLM applied in hindsight. All training items are excluded from evaluations to prevent contamination.
Helix runs entirely onboard embedded low-power-consumption GPUs, making it relevant for commercial deployment research and future humanoid applications. It uses a single set of neural network weights for all behaviors — picking and placing items, using drawers and refrigerators, and cross-robot interaction — without any task-specific fine-tuning. It has been demonstrated in household manipulation tasks and logistics package triaging, and can operate simultaneously on two robots through a supervisory architecture that decomposes overall goals into per-robot subtasks.
OpenVLA
OpenVLA is a 7B-parameter open-source VLA trained on a diverse collection of 970,000 real-world robot demonstrations. It builds on a Llama 2 language model combined with a visual encoder that fuses pretrained features from DINOv2 and SigLIP. Despite being 7× smaller, OpenVLA outperforms the closed RT-2-X (55B parameters) by 16.5 percentage points in absolute task success rate across 29 tasks and multiple robot embodiments.
A February 2025 paper introduced the OFT (Optimized Fine-Tuning) recipe, which combines parallel decoding, action chunking, a continuous action representation, and an L1 regression objective. OFT delivers 25–50x faster inference speed and achieves 97.1% average success rate on the LIBERO simulation benchmark, outperforming π0, Octo, and Diffusion Policy. An augmented version, OFT+, adds FiLM conditioning for better language grounding and enables high-frequency bimanual control on the ALOHA robot. OpenVLA supports LoRA fine-tuning and quantization for resource-constrained deployment, and community ROS 2 wrappers exist for integration with robot operating systems.
Octo
Octo is an open-source generalist robot policy from UC Berkeley, available in two sizes: Octo-Small (27M parameters) and Octo-Base (93M parameters). Both use a transformer backbone with diffusion decoding, pretrained on 800,000 robot episodes from the Open X-Embodiment dataset across 25 datasets. The model supports both natural language instructions and goal image conditioning, and accommodates flexible observation and action spaces including new sensors and action representations without architectural changes.
Octo was designed specifically to support efficient finetuning to new robot setups. In the official evaluation, each task uses approximately 100 target-domain demonstrations, and Octo outperforms training from scratch by an average of 52% across six evaluation setups spanning institutions including CMU, Stanford, and UC Berkeley. It performs comparably to RT-2-X (55B parameters) in zero-shot settings while being orders of magnitude smaller. Octo is primarily a research and developer tool, and is a strong lightweight starting point for labs needing to iterate quickly on new manipulation tasks with limited compute.
AGIBOT BFM and GCFM
In April 2026, Shanghai-based AGIBOT announced two foundation models as part of its “One Robotic Body, Three Intelligences” full-stack architecture. The Behavioral Foundation Model (BFM) is positioned around imitation and behavior transfer — designed to acquire new motion behaviors efficiently from demonstrations. The Generative Control Foundation Model (GCFM) is positioned around generating context-aware robot motions from multimodal inputs including text, audio, and video.
AGIBOT positions AGIBOT WORLD 2026 as part of the data foundation for its broader robotics stack — an open-source, production-grade real-world dataset spanning commercial spaces, homes, and everyday scenarios. The company declared 2026 its “Deployment Year One” at its April 2026 Partner Conference and announced the rollout of its 10,000th robot in March 2026.
Gemini Robotics On-Device
Gemini Robotics On-Device is a VLA model for bi-arm robots engineered to run locally on the robot itself with low-latency inference, without requiring a data network connection. Released in June 2025, it is the first VLA model Google DeepMind has made available for fine-tuning. It builds on the task generalization and dexterity capabilities of the cloud-based Gemini Robotics model, optimized for on-device execution where latency or connectivity constraints apply. The model was trained primarily on ALOHA robots and has been adapted to a bi-arm Franka FR3 and Apptronik’s Apollo humanoid. It adapts to new tasks with as few as 50 to 100 demonstrations. Availability is currently through selected trusted testers, not a general public release.
NVIDIA Cosmos World Foundation Models
Cosmos is not a robot policy model in the conventional sense — it is a generative world model that produces synthetic trajectory data to scale training pipelines for other models on this list. The GR00T-Dreams blueprint uses Cosmos to generate vast amounts of synthetic trajectory data from a single image and language instruction, enabling robots to learn new tasks in unfamiliar environments without requiring specific teleoperation data. This directly underpinned the development of GR00T N1.5. Cosmos Predict 2, the version used in GR00T-Dreams, is available on HuggingFace with performance enhancements for high-quality world generation and reduced hallucination. Companies including Skild AI and FieldAI are using Cosmos and Isaac simulation components to generate synthetic robot training data and validate robot behaviors in simulation before real-world deployment.
SmolVLA (HuggingFace LeRobot)
Released June 3, 2025, SmolVLA is HuggingFace’s 450M-parameter compact VLA built within the LeRobot framework and trained entirely on community-contributed open-source data. It uses a SmolVLM-2 vision-language backbone combined with a flow-matching transformer action expert — outputting continuous actions rather than discretized tokens, the same action representation used by π0 and GR00T N1. It was pretrained on 10 million frames curated from 487 community datasets tagged under “lerobot” on HuggingFace, spanning diverse environments from labs to living rooms.
SmolVLA runs on consumer hardware including single RTX-class GPUs and MacBooks. Official fine-tuning benchmarks show approximately 4 hours on a single A100 for 20,000 training steps. On real-robot evaluations using SO100 and SO101 arms, it achieves approximately 78.3% average success rate after task-specific fine-tuning. It matches or outperforms larger models like ACT on LIBERO and Meta-World simulation benchmarks, and supports asynchronous inference for 30% faster response and 2× task throughput. SmolVLA is the most accessible entry point into the VLA ecosystem for teams with limited compute.
