NVIDIA rolls out new Jetson Thor modules aimed at bringing humanoid and edge AI robots to scale

NVIDIA on Wednesday announced two new edge-computing modules under its Jetson Thor platform, positioning the devices as a path for robotics companies to move from prototypes to mass deployments that rely on “foundation model” intelligence at the device level. The new Jetson T3000 and T2000 are intended for general-purpose robots and autonomous machines that need high compute in a small power envelope, without requiring every decision to be processed in the cloud.

At the center of NVIDIA’s pitch is the Jetson AGX Thor platform, which the company said is already seeing growing adoption across industries and is being used as the hardware backbone for next-generation humanoids and robotic systems. NVIDIA named a broad set of companies building on the platform, including 1X, Agility, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, Medtronic and Techman Robot.

The Jetson T3000 module is aimed at performance-rich robotics and other edge AI applications. NVIDIA said it delivers 865 FP4 teraflops of AI compute in a compact form factor that is roughly half the size and power of the company’s earlier T5000 module. The company also said the T3000 pairs an NVIDIA Blackwell GPU with an eight-core Arm Neoverse CPU, 32GB of LPDDR5X memory, 273GB/s of memory bandwidth and 25 GbE connectivity. For safety-oriented robotics, NVIDIA also described an IGX T3000 variant that includes integrated functional safety and supports NVIDIA Halos for Robotics, a software stack intended to help robots operate alongside humans.

Despite the smaller footprint, NVIDIA said the T3000 can match the T5000’s inference performance for multimodal workloads. The company cited the kinds of models robots typically use, including large language models, vision-language models, vision-language-action models and “world foundation models.” NVIDIA added that migrating to the T3000 can reduce total costs, pointing to high memory prices as a driver for why lower-memory configurations may matter commercially.

The Jetson T2000 expands the Thor architecture to a wider range of edge AI systems. NVIDIA said it offers 400 FP4 teraflops of compute and 16GB of memory, targeting what it called an entry point for developers building visual AI agents, autonomous mobile robots, industrial manipulators and other intelligent machines.

NVIDIA also tied the new hardware to software tools meant to reduce friction for developers. The company said “Jetson agent skills” are newly released, designed to optimize the software stack and achieve meaningful memory savings in days rather than weeks. NVIDIA described the goal as helping developers run more capable workloads on lower-memory configurations, which it said can translate into lower system cost, faster deployment and the ability to move down one memory SKU within the same performance tier without giving up performance.

To illustrate the impact, NVIDIA pointed to case studies in which teams reduced memory usage enough to shift to smaller Jetson configurations. It cited humanoid robotics and related hardware: UBTech and Agile Robots, and an industrial solutions provider, Connect Tech, which NVIDIA said reduced memory usage by up to 15GB, enabling a move from a 64GB Jetson AGX Orin module to the 32GB Jetson configuration. In smart retail, SandStar was said to reduce memory usage by up to 4GB, enabling deployment on a Jetson Orin NX 8GB module instead of a 16GB configuration. For companion robotics, GROOVE X, maker of the LOVOT robot, was cited for using heterogeneous AI accelerators to optimize workload distribution and reduce memory. NVIDIA also cited intelligent transportation work at NoTraffic, which it said reduced memory usage by 30% on a Jetson TX2 NX platform.

Beyond memory optimization, NVIDIA said Jetson is “agentic-ready” for physical AI, highlighting the idea of on-device reasoning, autonomous decision-making and task automation. The company referenced NVIDIA NemoClaw blueprints for orchestrating intelligent agents, and it said the Jetson portfolio will support a broader path for developers building embodied AI systems.

A related software and model announcement accompanied the hardware. NVIDIA expanded its NVIDIA Cosmos 3 “frontier open world foundation model” family with Cosmos 3 Edge, described as a lightweight, 4-billion-parameter robot foundation model built for embodied systems. NVIDIA said Cosmos 3 Edge is compatible with NVIDIA Thor platforms and is intended to support real-time on-device inference for seeing the world, reasoning over it and predicting and generating actions, including the ability to post-train with specific embodiments and sensors using the open Cosmos framework in about a day. NVIDIA also said the new Jetson modules will run as part of the Thor family’s shared chip architecture and software stack, including tools such as NVIDIA Isaac for robotics simulation and perception. NVIDIA added that developers can begin using T3000 emulation mode later this month with JetPack 7.2.1, with T2000 emulation mode to follow in a future release, and that the Jetson T3000 and T2000 modules are scheduled for availability in Q1 2027.