RLDX-1 Technical Report

Summary

RLDX-1 is a VLA robot policy for dexterous manipulation that adds motion, memory, and tactile/torque inputs to a Qwen3-VL-based policy. The report claims large gains over π_{0.5} and GR00T N1.6 on simulation and real-robot tasks, especially high-DoF humanoid tasks.

Problem

Existing VLAs often handle scene understanding and language instructions, but they struggle when manipulation needs motion tracking, long-term task state, or contact sensing.
This matters for real robots because tasks such as conveyor-belt catching, shell-game-style selection, deformable-object grasping, and insertion can fail when the policy only sees the current image and instruction.

Approach

RLDX-1 uses Qwen3-VL 8B as the VLM base, fine-tuned with robot VQA data for spatial relations, subtask inference, and low-level action grounding.
The model adds a motion module for multi-frame video, a memory module that stores 3 past cognition-feature chunks, and a physics stream for tactile and torque signals.
Its Multi-Stream Action Transformer processes cognition, action/proprioception, and physics in separate streams, then lets them exchange information through joint self-attention while predicting action chunks with flow matching.
Training has 3 stages: multi-embodiment pretraining, embodiment-specific mid-training with in-house and synthetic data, and task-specific post-training with optional reinforcement learning.
The synthetic data pipeline uses image/video generation, inverse dynamics action labeling, video quality filtering, and motion-consistency filtering through simulator replay.

Results

On ALLEX humanoid tasks, RLDX-1 reports 86.8% success, while π_{0.5} and GR00T N1.6 are around 40%.
On conveyor-belt fast-object catching, RLDX-1 reports over 87.5% success, while π_{0.5} stays below 29.2%.
On GR-1 Tabletop, RLDX-1 reports 58.7% success versus 47.6% for GR00T N1.6.
On OpenArm versatile-intelligence tests, RLDX-1 improves Unseen Object success from 37.5% with π_{0.5} to 54.2%, and Unseen Task success from 45.8% to 54.2%.
On ALLEX Object-in-Box Selection, which needs long-term memory, RLDX-1 reports 91.7% success, while GR00T N1.6 and π_{0.5} are in the 30% range.
Synthetic data adds 9.1 percentage points on GR-1 Tabletop over real-only training, and inference optimization cuts per-step latency on RTX 5090 from 71.2 ms to 43.7 ms, a 1.63x speedup.

Link

https://arxiv.org/abs/2605.03269v2