Sold by: RLWRLD
Deployed on AWS
PLEASE SEARCH RLDX-1 IN "PROFESSIONAL SERVICE" FOR FURTHER IMPLEMENTATION
RLDX-1 is a dexterity-first robotics foundation model that unifies vision, language, touch, force, and memory in one architecture - giving robots human-level manipulation for grasping, pouring, and tool use.
Overview
PLEASE SEARCH RLDX-1 IN "PROFESSIONAL SERVICE" FOR FURTHER IMPLEMENTATION
RLDX-1 is a general-purpose robotics foundation model (RFM) built on a "Dexterity-First" philosophy - engineered from the ground up to give high-DoF robotic hands human-level dexterous manipulation. Unlike conventional Vision-Language-Action (VLA) models that reason primarily over what a robot sees and hears, RLDX-1 also processes torque, tactile feedback, and working memory within a single model. This is powered by its Multi-Stream Action Transformer (MSAT) architecture, which assigns independent streams to each modality - vision, language, action, force, and memory - and fuses them through joint attention, enabling robots to see, feel, remember, and adapt in dynamic, contact-rich environments. RLDX-1 delivers state-of-the-art performance across eight public robotics benchmarks. It scored 70.6 on RoboCasa Kitchen and 58.7 on GR-1 Tabletop-both surpassing NVIDIA's GR00T N1.6 - and reached 86.7% on LIBERO-Plus, which stress-tests robustness across lighting and camera variations. In real-world hardware testing on WIRobotics' ALLEX humanoid, RLDX-1 completed a contact-rich pot-to-cup pouring task with a 70.8% success rate, roughly double that of comparable models. Notably, it achieves this using approximately 20% of the training compute of NVIDIA GR00T N1.5, reflecting the strength of its architecture and data engine rather than sheer scale. Designed for real-world industrial deployment, RLDX-1 runs across single-arm, dual-arm, and humanoid embodiments with high-DoF hands, and has been validated on platforms including WIRobotics ALLEX, Franka Research 3, and OpenArm. It spans the complete robotics lifecycle - from a scalable data-collection pipeline combining teleoperation, synthetic, and human data, through training and optimized deployment - for tasks such as grasping, pouring, tool use, and assembly across manufacturing, logistics, retail, and service settings. Built on the NVIDIA robotics and AI stack, RLDX-1 is the foundation for dexterous automation at scale.
Highlights
- Dexterity-first, beyond vision and language. Unlike conventional Vision-Language-Action (VLA) models, RLDX-1 is a robotics foundation model that also processes torque, tactile feedback, and working memory in a single Multi-Stream Action Transformer (MSAT) - so robots don't just see and talk, but feel, remember, and adapt to master contact-rich dexterous manipulation like grasping, pouring, and tool use.
- State-of-the-art performance at a fraction of the compute. RLDX-1 leads eight public robotics benchmarks - including 70.6 on RoboCasa Kitchen and 86.7% on LIBERO-Plus - and doubled comparable models on a real-world pouring task (70.8% success). It reaches this dexterous manipulation performance using roughly 20% of the training compute of NVIDIA GR00T N1.5.
- One foundation model, every embodiment. RLDX-1 deploys across single-arm, dual-arm, and humanoid robots with high-DoF hands - validated on WIRobotics ALLEX, Franka Research 3, and OpenArm - covering the full lifecycle from data collection to optimized deployment for manufacturing, logistics, retail, and service automation.
Details
Sold by
Categories
Delivery method
Latest version
Deployed on AWS
New
Introducing multi-product solutions
You can now purchase comprehensive solutions tailored to use cases and industries.
Features and programs
Financing for AWS Marketplace purchases
AWS Marketplace now accepts line of credit payments through the PNC Vendor Finance program. This program is available to select AWS customers in the US, excluding NV, NC, ND, TN, & VT.
Pricing
Pricing is based on actual usage, with charges varying according to how much you consume. Subscriptions have no end date and may be canceled any time.
Additional AWS infrastructure costs may apply. Use the AWS Pricing Calculator to estimate your infrastructure costs.
Dimension | Description | Cost/host/hour |
|---|---|---|
ml.g5.xlarge Inference (Batch) Recommended | Model inference on the ml.g5.xlarge instance type, batch mode | $9,999,999.00 |
ml.g5.xlarge Inference (Real-Time) Recommended | Model inference on the ml.g5.xlarge instance type, real-time mode | $9,999,999.00 |
Vendor refund policy
No refunds will be made except
Legal
Vendor terms and conditions
Upon subscribing to this product, you must acknowledge and agree to the terms and conditions outlined in the vendor's End User License Agreement (EULA) .
Content disclaimer
Vendors are responsible for their product descriptions and other product content. AWS does not warrant that vendors' product descriptions or other product content are accurate, complete, reliable, current, or error-free.
Delivery details
Amazon SageMaker model
An Amazon SageMaker model package is a pre-trained machine learning model ready to use without additional training. Use the model package to create a model on Amazon SageMaker for real-time inference or batch processing. Amazon SageMaker is a fully managed platform for building, training, and deploying machine learning models at scale.
Deploy the model on Amazon SageMaker AI using the following options:
Version release notes
Version 1.0 - Initial Release This is the first release of RLDX-1, a dexterity-first robotics foundation model that unifies vision, language, action, tactile feedback, force sensing, and working memory in a single Multi-Stream Action Transformer (MSAT) architecture. What's included in v1.0 General-purpose dexterous manipulation across single-arm, dual-arm, and humanoid embodiments with high-DoF robotic hands. Multi-Stream Action Transformer (MSAT) architecture that fuses vision, language, action, force, and memory through joint attention for contact-rich tasks such as grasping, pouring, and tool use. State-of-the-art performance across eight public robotics benchmarks, including RoboCasa Kitchen and LIBERO-Plus. Validated deployment on platforms including WIRobotics ALLEX, Franka Research 3, and OpenArm. Built on the NVIDIA robotics and AI stack for simulation, training, inference, and deployment.
Additional details
Inputs
- Summary
RLDX-1 accepts a multimodal observation payload representing a single timestep of a robot's interaction with its environment. Powered by a Multi-Stream Action Transformer (MSAT) architecture, the model processes each modality as an independent input stream and fuses them through joint attention to produce dexterous manipulation actions. Inputs are submitted as a JSON request; image and sensor tensors are provided as [base64-encoded arrays / referenced tensors] in the format below. Input components Visual observation - One or more RGB camera frames capturing the scene and the robot's workspace (e.g., a head/front camera and, optionally, wrist-mounted cameras). Format: [H × W × 3] uint8 images at [resolution, e.g., 224×224]. Multiple camera views may be supplied as a list keyed by camera name. Language instruction - A natural-language task instruction as a UTF-8 text string (e.g., "pick up the cup and pour water into the bowl"). This conditions the model on the intended task and target objects. Robot state (proprioception) - The robot's current proprioceptive state, including joint positions and, where applicable, end-effector pose and gripper/hand joint angles for high-DoF hands. Format: float array of dimension [robot-specific DoF]. Force / torque signals - Force-torque and/or joint-torque readings that let the model perceive contact and load during contact-rich tasks such as grasping and pouring. Format: float array [sensor-specific dimension]. [Optional depending on embodiment.] Tactile signals - Tactile sensor readings from the fingertips/hand surfaces, used for fine in-hand manipulation. Format: [sensor-specific array]. [Optional depending on embodiment.] Embodiment / configuration metadata - Identifier for the target robot embodiment (e.g., single-arm, dual-arm, or humanoid with high-DoF hands) so the model applies the correct action space. [Specify accepted embodiment identifiers.] The model maintains working memory internally across successive inference calls, so requests should be submitted sequentially in temporal order to preserve context for long-horizon tasks.
- Input MIME type
- application/x-msgpack
{
"instruction": "Pick up the cup and pour water into the bowl",
"embodiment": "dual_arm_humanoid",
"images": {
"head_camera": "<base64-encoded RGB image, 224x224x3, uint8>",
"left_wrist_camera": "<base64-encoded RGB image, 224x224x3, uint8>",
"right_wrist_camera": "<base64-encoded RGB image, 224x224x3, uint8>"
},
"robot_state": {
"joint_positions": [0.12, -0.45, 0.88, -1.20, 0.03, 0.57, -0.31,
0.10, -0.42, 0.90, -1.18, 0.05, 0.55, -0.29],
"end_effector_pose": {
"left": [0.31, 0.18, 0.52, 0.0, 0.0, 0.0, 1.0],
"right": [0.29, -0.17, 0.53, 0.0, 0.0, 0.0, 1.0]
},
"hand_joint_angles": {
"left": [0.05, 0.12, 0.08, 0.10, 0.06, 0.11, 0.09, 0.13, 0.07, 0.10, 0.08, 0.12],
"right": [0.04, 0.11, 0.07, 0.09, 0.05, 0.10, 0.08, 0.12, 0.06, 0.09, 0.07, 0.11]
}
},
"force_torque": {
"left": [1.2, -0.4, 3.8, 0.05, -0.02, 0.01],
"right": [1.0, -0.3, 3.6, 0.04, -0.01, 0.02]
},
"tactile": {
"left_fingertips": [[0.02, 0.05, 0.01], [0.03, 0.04, 0.02], [0.01, 0.06, 0.03], [0.02, 0.05, 0.02], [0.03, 0.04, 0.01]],
"right_fingertips": [[0.01, 0.04, 0.02], [0.02, 0.05, 0.01], [0.03, 0.04, 0.02], [0.01, 0.06, 0.03], [0.02, 0.05, 0.02]]
},
"timestep": 0
}
{"id": "episode_001_t0000", "instruction": "Pick up the cup and pour water into the bowl", "embodiment": "dual_arm_humanoid", "images": {"head_camera": "<base64 RGB 224x224x3 uint8>", "left_wrist_camera": "<base64 RGB 224x224x3 uint8>", "right_wrist_camera": "<base64 RGB 224x224x3 uint8>"}, "robot_state": {"joint_positions": [0.12, -0.45, 0.88, -1.20, 0.03, 0.57, -0.31, 0.10, -0.42, 0.90, -1.18, 0.05, 0.55, -0.29], "hand_joint_angles": {"left": [0.05, 0.12, 0.08, 0.10, 0.06, 0.11, 0.09, 0.13, 0.07, 0.10, 0.08, 0.12], "right": [0.04, 0.11, 0.07, 0.09, 0.05, 0.10, 0.08, 0.12, 0.06, 0.09, 0.07, 0.11]}}, "force_torque": {"left": [1.2, -0.4, 3.8, 0.05, -0.02, 0.01], "right": [1.0, -0.3, 3.6, 0.04, -0.01, 0.02]}, "tactile": {"left_fingertips": [[0.02, 0.05, 0.01], [0.03, 0.04, 0.02], [0.01, 0.06, 0.03], [0.02, 0.05, 0.02], [0.03, 0.04, 0.01]], "right_fingertips": [[0.01, 0.04, 0.02], [0.02, 0.05, 0.01], [0.03, 0.04, 0.02], [0.01, 0.06, 0.03], [0.02, 0.05, 0.02]]}, "timestep": 0}
{"id": "episode_001_t0001", "instruction": "Pick up the cup and pour water into the bowl", "embodiment": "dual_arm_humanoid", "images": {"head_camera": "<base64 RGB 224x224x3 uint8>", "left_wrist_camera": "<base64 RGB 224x224x3 uint8>", "right_wrist_camera": "<base64 RGB 224x224x3 uint8>"}, "robot_state": {"joint_positions": [0.13, -0.44, 0.87, -1.19, 0.04, 0.58, -0.30, 0.11, -0.41, 0.89, -1.17, 0.06, 0.56, -0.28], "hand_joint_angles": {"left": [0.06, 0.13, 0.09, 0.11, 0.07, 0.12, 0.10, 0.14, 0.08, 0.11, 0.09, 0.13], "right": [0.05, 0.12, 0.08, 0.10, 0.06, 0.11, 0.09, 0.13, 0.07, 0.10, 0.08, 0.12]}}, "force_torque": {"left": [1.3, -0.5, 4.0, 0.06, -0.03, 0.02], "right": [1.1, -0.4, 3.7, 0.05, -0.02, 0.01]}, "tactile": {"left_fingertips": [[0.03, 0.06, 0.02], [0.04, 0.05, 0.03], [0.02, 0.07, 0.04], [0.03, 0.06, 0.03], [0.04, 0.05, 0.02]], "right_fingertips": [[0.02, 0.05, 0.03], [0.03, 0.06, 0.02], [0.04, 0.05, 0.03], [0.02, 0.07, 0.04], [0.03, 0.06, 0.03]]}, "timestep": 1}
{"id": "episode_002_t0000", "instruction": "Grasp the tool and place it on the tray", "embodiment": "single_arm", "images": {"head_camera": "<base64 RGB 224x224x3 uint8>", "wrist_camera": "<base64 RGB 224x224x3 uint8>"}, "robot_state": {"joint_positions": [0.20, -0.30, 0.75, -1.05, 0.10, 0.60, -0.25], "hand_joint_angles": {"main": [0.08, 0.14, 0.10, 0.12, 0.09, 0.13, 0.11, 0.15, 0.10, 0.12, 0.10, 0.14]}}, "force_torque": {"main": [0.9, -0.2, 3.2, 0.03, -0.01, 0.01]}, "tactile": {"fingertips": [[0.02, 0.04, 0.02], [0.03, 0.05, 0.02], [0.01, 0.06, 0.03], [0.02, 0.05, 0.02], [0.03, 0.04, 0.01]]}, "timestep": 0}
Support
Vendor support
RLWRLD provides dedicated support to help you deploy and scale RLDX 1 across your robotics fleet. Our team assists with onboarding, model integration across single arm, dual arm, and humanoid embodiments, fine tuning for your target tasks, and deployment optimization. Contact Email: [operation@rlwrld.ai ] Support portal / docs: [https://www.rlwrld.ai ] What buyers can expect Onboarding support to get RLDX 1 running on your hardware and simulation environment, including guidance on the NVIDIA robotics and AI stack. Technical assistance covering model integration, fine tuning, tactile/force sensor setup, and deployment troubleshooting. Access to technical documentation, integration guides, and model resources. Initial response within [one business day] for standard inquiries during business hours ([KST 09:00 18:00, Mon Fri]), with priority handling for deployment critical issues. For enterprise engagements, custom fine tuning, or joint proof of concept projects, our solutions team offers dedicated onboarding and a named technical point of contact. To discuss enterprise support tiers or partnership opportunities, please reach out to [operation@rlwrld.ai ].
AWS infrastructure support
AWS Support is a one-on-one, fast-response support channel that is staffed 24x7x365 with experienced and technical support engineers. The service helps customers of all sizes and technical abilities to successfully utilize the products and features provided by Amazon Web Services.
Customer reviews
No customer reviews yet
Be the first to review this product . We've partnered with PeerSpot to gather customer feedback. You can share your experience by writing or recording a review, or scheduling a call with a PeerSpot analyst.