To build a streaming robot policy with NVIDIA Cosmos3-DROID, stage the DROID demonstrations in the expected LeRobotDataset v3.0 layout, convert a compatible Cosmos checkpoint to PyTorch Distributed Checkpoint (DCP), filter the training windows, post-train the action policy, and connect a serving process to a robot client. The documented Nano recipe is a worked example—not a plug-and-play configuration for every robot—and its reproduction run disables evaluation. Training, inference, and closed-loop evaluation are separate parts of the pipeline.
What the Cosmos3-DROID pipeline produces
The NVIDIA Cosmos Framework recipe post-trains Cosmos3-Nano to predict robot actions from video observations and proprioceptive state. Its reference policy uses 480p observations with concatenated camera views and predicts chunks of 32 future actions. The action representation is 8-dimensional, including the gripper, and specifies absolute joint positions. These are recipe-specific settings, not universal defaults for other embodiments. NVIDIA’s DROID action-policy post-training recipe documents the setup.
As an Amazon Associate I earn from qualifying purchases.
The data behind the recipe is the NVIDIA Cosmos3-DROID release. Its 2026 dataset card reports 76,000 teleoperated trajectories and approximately 350 hours of interaction data across 86 tasks and 564 scenes. It describes contributions from 50 data collectors across 18 labs and 13 institutions. Episodes include synchronized stereo RGB camera streams, calibration and depth information, robot state, control commands, and up to three natural-language instructions.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
The card identifies the collection platform as a Franka Panda 7-DoF arm with a Robotiq 2F-85 gripper. That hardware context matters: the reference action and observation setup reflects a particular robot and sensor arrangement. A new embodiment needs its own compatible action definition, state inputs, camera mapping, and normalization; changing the robot is not just a matter of pointing the loader at different videos.
#1 Best Overall
- The NVIDIA Jetson AGX Orin 64GB Developer Kit makes it easy to get started with Jetson Orin. Compact size, lots of connectors, and up to 275 TOPS of AI performance make this developer kit perfect for prototyping advanced AI-powered robots and other autonomous machines.
- The developer kit includes a Jetson AGX Orin 64GB module, and can emulate all the Jetson Orin modules. It supports multiple concurrent AI application pipelines with the NVIDIA Ampere GPU architecture, next-generation deep learning and vision accelerators, high-speed IO and fast memory bandwidth. Now you can develop solutions using your largest and most complex AI models to solve problems such as natural language understanding, 3D perception, and multi-sensor fusion.
- Jetson runs the NVIDIA AI software stack, and use-case specific application frameworks are available, including Isaac for robotics, DeepStream for vision AI, and Riva for conversational AI. You can save significant time with NVIDIA Omniverse Replicator for synthetic data generation (SDG), and by using NVIDIA TAO toolkit to fine-tune pretrained AI models from the NGC catalog.
- Jetson ecosystem partners offer additional AI and system software, developer tools, and custom software development. They can also help with cameras and other sensors, as well as carrier boards and design services for your product.
- With the computing capability of more than 8 Jetson AGX Xavier systems in a developer kit that integrates the latest NVIDIA GPU technology with the world’s most advanced deep learning software stack, you’ll have the flexibility to create tomorrow’s AI solution as well as today’s.
How to post-train Cosmos 3 on DROID data
- Stage the dataset. Download the
nvidia/Cosmos3-DROIDrelease and arrange it in the directory layout expected by the recipe’s loader. The recipe expects LeRobotDataset v3.0 format and a pre-downloaded dataset. - Convert the base checkpoint. Convert the selected Cosmos base checkpoint to PyTorch Distributed Checkpoint (DCP), the format expected by this post-training workflow.
- Apply the curation filter. Use
keep_ranges_1_0_1.jsonto omit idle or non-task time windows. The maintained recipe describes the retained window set as approximately 74% of windows; this is a curation proportion, not a claim that 74% of trajectories succeed or that the model reaches a particular accuracy. - Launch the registered experiment. Run the DROID action-policy experiment and save checkpoints. The maintained Nano recipe uses HSDP and is designed for one node with eight GPUs or a larger multi-node run. Its documented settings include global batch size 8192, learning rate
2e-4, and action chunk length 32. - Export or serve the trained policy. Connect a client that supplies the expected observation dictionary and receives action chunks. Serving is a separate deployment step from post-training.
The recipe’s documented reproduction run has evaluation disabled. Completing that run can produce a trained checkpoint, but the run itself does not establish policy performance.
How policy-server streaming works
The policy server accepts an observation dictionary from a client and returns an action chunk. NVIDIA documents server configurations for the Nano and Edge DROID variants, as well as a RoboLab simulation client, in its Cosmos3-Policy-DROID server guide. In practice, the client and policy must agree on the observation fields, tensor or data formats, camera arrangement, and action interpretation; a successful network connection alone does not make an incompatible robot interface safe to execute.
Rank #2
- Complete Jetson Orin Nano Starter Kit: This jetson orin nano starter kit includes a 30-in-1 sensor board, 8MP camera, dual-servo gimbal, 128GB SD card, and essential accessories. It supports Avisual recognition and voice interaction, providing a complete AI application development experience
- 8MP AI Vision Camera with Gimbal: Equipped with an IMX219 8MP camera and dual-servo gimbal, the jetson orin nano development kit supports face tracking, object recognition, target tracking, and computer vision projects. Ideal for learning AI vision, edge computing, robotics, and intelligent automation applications
- 11.6-Inch HD Display & AI Voice Assistant: Features an 11.6-inch 1366×768 IPS screen, allowing users to develop and test projects without an external monitor. The built-in AI voice interaction system supports voice commands and intelligent conversations, creating a more engaging and interactive learning experience
- 30 Sensors and 38 Guided Python Tutorials: Features a 30-in-1 sensor board with temperature & humidity, ultrasonic ranging, gas, motion, and other commonly used sensors. Includes 38 guided Python tutorials covering sensor applications, embedded development, and AI visual recognition from beginner to advanced
- Portable All-in-One Design with Rich Expansion Options: The Jetson Orin Nano Dev Kit provides multiple expansion interfaces including I2C/UART/IO interfaces. A custom carrying case integrates all components, making it convenient for classroom teaching, laboratory projects, demonstrations, and mobile AI development
Streaming is chunk-based rather than a fresh prediction for every camera frame. NVIDIA’s Edge tutorial describes the policy as replanning after each inference cycle, not after each observation. The client executes the current chunk while the next inference cycle prepares another, allowing motion to continue between predictions. This makes inference time, communication, and robot execution part of one control loop.
Can Cosmos 3 Edge run a robot policy on Jetson Thor?
NVIDIA’s August 19, 2026 tutorial adapts the action-policy recipe to Cosmos3-Edge and demonstrates on-device inference on a Jetson AGX Thor T5000. The tutorial reports approximately 1.53 seconds to generate a chunk, which covers roughly 2.13 seconds of robot motion in that setup. NVIDIA says the next chunk is prepared before the current motion finishes. These figures describe the tutorial’s hardware and configuration; they are not general latency guarantees, and the reported inference cycle does not mean the robot replans after every observation. NVIDIA’s Edge tutorial gives the deployment details.
Rank #3
- Supercharged AI Performance: Powered by NVIDIA Jetson Orin NX 16GB, delivers up to 157 TOPS in MAXN Super Mode — ideal for vision AI, robotics, autonomous machines, and generative AI workloads.
- Advanced Thermal Engineering for Full-Power Operation: Equipped with a vacuum copper heat pipe system, ultra-low thermal resistance medium, and high-emissivity black-coated surface combined with high-performance active cooling — ensuring stable full compute power even at 60°C ambient temperature.
- Energy-Efficient & Flexible Power Modes: Adjustable power profile from 10W to 40W, enabling a perfect balance between performance and efficiency for edge AI computing in diverse environments.
- Industrial-Grade Reliability & Design: Ruggedized for operation from -20°C to 60°C at 40W (up to 65°C at 25W), providing dependable performance in industrial automation and outdoor AI deployments.
- Rich Connectivity & AI-Ready Platform: Features 2×RJ45, SIM slot, 4×USB 3.2, HDMI 2.1, CAN, M.2 Key E/M, Mini-PCIe, and 4×CSI camera ports — supporting multi-camera vision, IoT, and robotics projects. Pre-installed with JetPack 6.2 and 128GB NVMe SSD, fully compatible with NVIDIA Isaac, ROS 1/2, and Hugging Face frameworks.
The tutorial reports 22.9% success across 120 language-conditioned manipulation tasks in closed-loop RoboLab evaluation. This is NVIDIA’s result for that simulated evaluation context, not a general real-world success rate or evidence that the same outcome will transfer to another robot, task set, or deployment.
For the tutorial’s Edge training path, NVIDIA lists DGX Station configurations with GB200 or GB300 systems as validated training hardware and describes a large multi-node training run. The tutorial’s duration details are inconsistent: its prerequisites refer to 60,000 iterations and roughly 68 hours, while its configuration table describes a 10,000-iteration run. Treat the exact training duration as unresolved unless the recipe or model card clarifies it. Jetson Thor is the tutorial’s inference target, not the training system for that multi-node job.
Rank #4
- 【Core Parameters】★AI Perf: 34/67 TOPS ★GPU:1024-core official Ampere architecture GPU with 32 Tensor Cores ★CPU:6-core Arm Corte-A78AE v8.2 64-bit CPU 1.5MB L2 + 4MB L3 ★Memory:8GB 128-bit LPDDR5 68 GB/s ★Storage: external NVMe via M.2 Key M
- 【Empowered by Large Al Model, Enhanced Human-Computer Interaction】Jetson Orin Super leverages three AI models and incorporates an AI voice interaction module. This multimodal visual system matches the scene being described, enabling environmental awareness and AI visual gameplay. Combined with a large-scale voice module and camera, it enables speech-to-text, semantic analysis, natural conversation, and real-time video analysis, enabling advanced embodied AI applications.
- 【AI Upgrade】Jetson Orin Nano series modules are compact in size but can deliver up to 34-67 TOPS of AI performance, with power consumption ranging from 7 watts to 25 watts. Compared to the Jetson Nano B01, it offers up to 80 times the performance and sets a new standard for entry-level edge AI.
- 【Highly compatible carrier board】Yahboom's carrier board is fully compatible with orin nano module. Compared to carrier boards that use Jetson Nano on the market, the newly upgraded circuit supports 25W power mode, which enables larger and more complex neural networks and fully leverages the performance of the core module. The resources, size, and interfaces of the Yahboom carrier board are consistent with the official board, with the only difference addition of power switch button.
- 【Tutorial materials provided】The JETSON system based on Ubuntu 22.04 provides a complete desktop Linux environment with accelerated graphics, supporting CUDA 12.6, TensorRT 10.7.0, cuDNN 9.6.0, OpenCV 4.10.0, etc. The performance on AI LLM, VLM and visual Transformer is significantly improved compared with the previous generation.
Choosing a Cosmos model and deployment path
NVIDIA’s model reference distinguishes the generator, used for tasks such as world generation, simulation, future prediction, synthetic data generation, and policy learning, from the reasoner, used for world understanding, grounding, planning, and decision-making. The documented DROID pathways center on generator models.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →| Model | Parameters | Documented fit in this workflow |
|---|---|---|
| Cosmos3-Super | 64B | High-quality generation and synthetic data work; not the base model in the main DROID post-training recipe. |
| Cosmos3-Nano | 16B | Base model used by the maintained DROID action-policy post-training recipe; intended as a balanced post-training option. |
| Cosmos3-Edge | 4B | Compact model demonstrated for edge deployment and on-device policy inference. |
Model sizes and descriptions are from NVIDIA’s Cosmos model reference; the workflow positioning is also described in the Cosmos repository. Selecting a model is only one part of deployment planning. Compare the available GPU memory and count, node capacity and training time, where inference will run, end-to-end control latency, and the effort required to match the policy to the robot’s embodiment. Evaluation evidence should also be judged by whether it comes from simulation or a physical robot, which tasks it covers, how success is defined, and whether the setup is reproducible.
What must change for another robot
The DROID recipe should be treated as a reference implementation for its data and embodiment, not as a generic controller. Before adapting it, define and validate the robot-specific interface:
- Actions: Set the action dimensions, joint ordering, gripper representation, and whether commands are absolute positions or another control format.
- State: Map the robot’s proprioceptive signals to the policy’s expected inputs and verify their units and normalization.
- Cameras: Specify which camera streams are used, how views are ordered or concatenated, and how image preprocessing matches training.
- Instructions and timing: Align task-instruction fields and observation timing with the client and server interface, then check that chunk execution and replanning cadence fit the robot.
- Evaluation: Validate in an appropriate closed-loop environment before relying on a policy in physical operation. Results from RoboLab simulation do not by themselves establish safe or effective behavior on hardware.
These adaptations are essential because a policy trained against one action space and sensor layout can produce commands that are syntactically valid yet physically wrong for a different robot.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →
