RLDX-1

RLDX-1 is a dexterity-first foundation model for robotic hands developed by RLWRLD and introduced in May 2026. The model uses the proprietary Multi-Stream Action Transformer (MSAT) architecture, in which each modality (vision, language, proprioception, memory, tactile, torque) is processed in its own dedicated stream and joint self-attention fuses them before action decoding.

RLDX-1 uses a fine-tuned Qwen3-VL 8B as its vision-language backbone (RLDX-1-VLM). The model integrates a Motion Module (multi-frame compression through cognition video tokens), a Physics Module (tactile and torque stream with future-contact-state prediction) and a Cognition Interface with 64 cognition tokens that doubles as the substrate for the long-horizon Memory Module.

RLDX-1 ships in three checkpoints: RLDX-1-PT (pre-trained, embodiment-agnostic) and two 8.1B mid-trained variants — RLDX-1-MT-ALLEX (ALLEX humanoid) and RLDX-1-MT-DROID (Franka Research 3 with AnySkin tactile). The training pipeline includes pre-training, mid-training for the target embodiment, and post-training with DAgger plus Progress-Aware RL driven by a dedicated VLM-critic. Weights and code are released on Hugging Face and GitHub.

In simulation, RLDX-1 reaches 97.8 on LIBERO, 70.6 on RoboCasa Kitchen, 58.7 on RoboCasa GR-1 Tabletop and 32.1 on RoboCasa 365, outperforming π₀.₅, π₀-FAST and GR00T N1.5/N1.6. On the real ALLEX benchmark (Conveyor Pick-and-Place, Object-in-Box Selection, Pot-to-Cup Pouring) RLDX-1-MT-ALLEX scores 87.5%, 91.7% and 70.8% respectively, versus below 30% for the baselines.