RLWRLD presents RLDX-1, a dexterity-first foundation model for high-degree-of-freedom robotic hands

RLWRLD last week presented RLDX-1, designed to tackle complex tasks in real-world industry using robotic hands. The model integrates a scalable data-collection pipeline, versatile architecture design, robust training methodologies, and optimized deployment strategies, and is deployable across single-arm, dual-arm, and humanoid embodiments.

RLDX-1 uses a Multi-Stream Action Transformer (MSAT) architecture where each sensory modality—torque (a high-rate continuous stream), video (sparse high-dimensional frames), and memory (stateful)—gets its own dedicated processing stream. The streams communicate through joint self-attention without being forced into a shared representation prematurely. A robot-specialized vision language model fine-tuned on robot visual question and answering targets spatial reasoning, task understanding, and action grounding.

On RoboCasa, the fine-tuned vision model achieved +3.42 percentage points over the vanilla model. On conveyor-belt pick-and-place tasks, the Motion Module achieved +37.5 percentage points over GR00T N1.6 and π₀.₅. The Cognition Interface achieved +35 percentage points inference speedup (16.3→22.1 Hz).