Agile ONE SN

The Agile ONE SN represents a major shift in physical AI and embodied agent hardware. Built by Agile Robots as a humanoid reference platform powered by the Nvidia Jetson Thor system-on-chip, the robot is designed to bridge the gap between digital AI agents and physical task execution. Unlike static edge servers or wheeled mobile manipulators, the Agile ONE SN packs 128GB of unified memory and specialized AI compute directly inside an untethered, 1.4-meter-tall bipedal frame. It targets enterprise R&D departments, academic robotics labs, and teams developing autonomous agentic workflows that must operate directly in human environments.

Positioned in the specialized physical AI research tier, the platform serves as an alternative to piecing together discrete compute boxes, separate robotic arms, and third-party vision sensors. When engineering teams evaluate unknown hardware for AI development, integration friction often burns months before a model can execute an action in the real world. The Agile ONE SN tackles this by combining high-density onboard compute with pre-calibrated sensor suites and native integration into Nvidia's Isaac GR00T 1.7 ecosystem.

For developers looking for the best hardware for running AI models locally in an embodied form factor, this machine treats large language models, vision-language models (VLMs), and vision-language-action (VLA) policies as first-class citizens. By integrating compute at the edge with tactile five-finger hands, the hardware enables zero-latency closed-loop motor control combined with high-level cognitive reasoning on a single unified memory pool.

The compute engine behind the Agile ONE SN is Nvidia's Jetson Thor module. Designed specifically to serve as the best AI chip for local deployment in robotics, Thor provides the mathematical throughput required for modern generative AI and sensor fusion:

• Compute Throughput: Up to 2,070 FP4 TFLOPS (also supporting FP8, INT8, and FP16 precisions).
• Onboard Memory: 128GB unified memory, acting effectively as an ultra-high-capacity 128GB GPU for AI workloads.
• Physical Envelope: 1.4 m (4.59 ft) height, 43 kg (95 lb) curb weight.
• Degrees of Freedom (DoF): 71 total DoF (31 body joints, 20 DoF per hand).
• Manipulators: Dual Wuji five-finger hands with tactile sensing and direct force control.
• Payload Capacity: Up to 3 kg (6.6 lbs) per arm.
• Sensor Array: Head-mounted RGB-D cameras, dual wrist-mounted cameras, 3D lidar, internal IMUs/motion sensors.
• Connectivity: Wi-Fi 7, Bluetooth, hardware emergency stop.
• Software Stack: Nvidia Isaac GR00T 1.7, Isaac Sim, Isaac Lab.

The defining specification for Agile ONE SN AI inference performance is its 128GB unified memory architecture. In typical edge robotics, perception runs on one low-power GPU while high-level planning must be offloaded via Wi-Fi to a rack server. The Agile ONE SN avoids network latency bottlenecks by sharing 128GB of high-bandwidth memory between the CPU cores and the Blackwell-generation GPU compute units on Jetson Thor. This shared memory pool allows raw lidar and stereo camera frames to stream directly into vision models without redundant PCIe copy operations, freeing memory bus capacity for transformer token generation.

Thermal and power constraints in a 43 kg bipedal chassis require strict efficiency. Jetson Thor achieves its 2,070 TFLOPS rating using native FP4 tensor cores, delivering high inferencing throughput per watt. This efficiency ensures that executing multi-billion-parameter foundation models does not immediately exhaust the onboard battery system or cause thermal throttling during dynamic motor control loops.

The generous Agile ONE SN VRAM for large language models eliminates the memory ceiling that historically crippled humanoid robotics. With 128GB of addressable memory, the platform operates comfortably as a self-contained local LLM and VLA inferencing node.

The platform serves as viable mobile hardware for running 70B parameter models entirely on device:

• Llama 3.1 70B: Quantized to INT4 or FP4, the weights occupy roughly 35GB to 40GB of VRAM. This leaves more than 80GB of headroom for real-time vision processing, system memory, and motion-planning buffers.
• Qwen 2.5 72B: At 4-bit precision (AWQ or GPTQ), Qwen runs smoothly inside a 42GB footprint, providing advanced multi-step reasoning directly onboard.
• DeepSeek-R1 Distillations: Distillations like DeepSeek-R1-Distill-Llama-70B can fit easily with large context windows for complex physical environment parsing.

• Isaac GR00T Foundation Models: Built explicitly to pair with Nvidia Isaac GR00T 1.7, the platform runs GR00T VLA architectures natively. These models take multimodal visual and tactile inputs and produce 71-axis joint trajectories at high control frequencies.
• Pixtral 12B, Llama 3.2 Vision (11B), and Qwen 2-VL (7B): These models fit comfortably in unquantized 16-bit or 8-bit formats, processing head and wrist camera frames locally for real-time spatial grounding.

Agile ONE SN tokens per second will vary based on concurrent motor-control workloads:

• For 8B models (such as Llama 3.1 8B at FP16 or INT8), expect swift generation speeds of 40 to 60+ tokens per second, ideal for interactive conversational robotics.
• For 70B models running at FP4/INT4 precision, the system balances generation at an estimated 12 to 20 tokens per second. This speed is fast enough for high-level agentic task orchestration without stalling physical manipulation routines.
• The Sweet Spot: The ideal quantization setup for this system is FP8 for vision-action policies (preserving the subtle tactile gradients from the Wuji hands) paired with 4-bit (INT4/FP4) weights for background reasoning LLMs.

The Agile ONE SN for AI development is targeted at specialized teams deploying autonomous agents into dynamic environments:

• Embodied AI and Foundation Model Researchers: Labs testing whether internet-scale pretraining translates into physical dexterity can run end-to-end VLA models without tethering the robot to an umbilical cable or local server rack.
• Industrial Automation Teams: With 3 kg payload per arm and 20 degrees of freedom per hand, industrial engineers can test fine-motor assembly, bin picking, and tool use where force feedback and vision must loop continuously.
• Developers of Local AI Agents (2026): For practitioners evaluating the best hardware for local AI agents 2026 has to offer, the Agile ONE SN represents a production-grade embodiment platform. It allows software agents to leave the browser and navigate real office or factory spaces.
• Edge Inference Testing: Teams testing the limitations of running an Agile ONE SN local LLM stack can benchmark real-world battery drain, thermal behavior, and cognitive latency in unstructured physical spaces.

This is not a toy for casual hobbyists running basic local chatbots. It is an industrial-grade research instrument intended for engineers writing custom ROS 2 nodes, Isaac Lab reinforcement learning pipelines, and low-level TensorRT-LLM inference engines.

When evaluating Agile ONE SN vs Unitree G1, the hardware designs reflect distinctly different operational philosophies:

• Unitree G1: The G1 is an ultra-compact, budget-oriented research humanoid that relies on substantially lighter onboard compute (typically low-power embedded modules). Running large models on the G1 often requires streaming sensor data offboard to an external PC or GPU server, introducing network latency and Wi-Fi drop vulnerabilities.
• Agile ONE SN: Agile Robots chose a compute-first approach. By embedding the Nvidia Jetson Thor with 128GB of memory and up to 2,070 FP4 TFLOPS directly into the torso, the Agile ONE SN functions as an independent, fully autonomous edge computing server on legs.

Furthermore, while competitors often use simplified 3-finger grippers, the Agile ONE SN provides dual 20-DoF Wuji tactile hands with active force control. If your research workload demands high-parameter onboard reasoning, 70B model execution without cloud reliance, and delicate physical object manipulation, the Agile ONE SN provides an integrated, deployment-ready foundation.