Description
RLDX-1 is a dexterity-first foundation model from RLWRLD for robots with high-degree-of-freedom hands. It targets tasks that need fine motor control, such as pouring a drink as the container gets lighter, rotating an object in hand, or grasping something moving on a conveyor, cases where typical vision-language-action (VLA) models fail because they were never given the right signals: hand-object geometry, contact force, and timing.
Read More
The model uses an architecture RLWRLD calls MSAT (Multi-Stream Action Transformer), which gives each modality, vision, motion, torque and tactile signals, and memory, its own processing stream instead of forcing everything through one transformer. This structure was built to address five capability regimes from RLWRLD's own DexBench benchmark: grasp diversity, spatial precision, temporal precision, contact precision, and context awareness. Specific components handle each regime, including a robot-specialized vision-language model for spatial reasoning, a Motion Module for tracking fast-moving objects, a Physics Module for torque and tactile sensing, and a memory-and-recovery system for multi-step tasks.
RLDX-1 is trained on synthetic robot data (scaled roughly 5x from a small teleoperation set) plus human demonstration data captured from bare human hands and retargeted onto robot hands, yielding more than 200 demonstrations per hour. Training runs through three stages: pre-training for general manipulation across embodiments, mid-training on target platforms including an ALLEX humanoid and a Franka Research 3 arm, and post-training that uses DAgger-based recovery and progress-aware reinforcement learning, which RLWRLD reports lets the final policy complete tasks about 3x faster than imitation learning alone.
RLDX-1 ships as three 8.1-billion-parameter checkpoints, a general pre-trained model plus two mid-trained versions for the ALLEX and DROID platforms, with public checkpoints on Hugging Face, code on GitHub, and a technical report on arXiv. RLWRLD benchmarks it against models including GR00T N1.5/N1.6 and the pi-0 family in simulation, and on the OpenArm real-world benchmark, and reports inference speedups of up to 1.63x through static-graph compilation and operator fusion. It's aimed at robotics researchers and industrial manufacturers working on dexterous manipulation rather than at end users.
- CategoryRobotics & AI Hardware
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
RLDX-1 Highlights
Dexterity-first foundation model for high-DoF robot hands
Multi-Stream Action Transformer (MSAT) architecture with a separate stream per modality
Robot-specialized vision-language model for spatial precision
Motion Module for tracking fast-moving objects
Physics Module for torque and tactile contact sensing
Memory and DAgger-based recovery for multi-step task context
Trained on synthetic and human-retargeted demonstration data
Public checkpoints on Hugging Face, code on GitHub, and a technical report on arXiv
Works across single-arm, dual-arm, and humanoid embodiments
Getting Started with RLDX-1
Read the report: Review RLWRLD's technical report and results on arXiv.
Get the checkpoints: Download the RLDX-1 checkpoints from RLWRLD's Hugging Face collection.
Get the code: Clone the RLDX-1 repository on GitHub.
Pick a checkpoint: Start from RLDX-1-PT, or the ALLEX or DROID mid-trained versions.
Fine-tune for your embodiment: Adapt the model to a specific robotic hand or platform.
Benchmark your setup: Compare results against RLWRLD's published DexBench and OpenArm figures.
RLDX-1's Use Cases
- Precision manipulation
- Industrial dexterity
- Humanoid robot research
- Dynamic object handling
