RLDX-1 splits motion, memory and touch into separate streams: 86.8% average success on ALLEX humanoid tasks, against about 40% for π0.5 and GR00T N1.6
Synopsis
RLWRLD and KAIST released RLDX-1, a general-purpose robot policy that adds motion awareness, long-term memory and physical sensing on top of a vision-language-action model. It keeps each modality in its own stream inside a Multi-Stream Action Transformer and lets the streams exchange information through joint self-attention, paired with synthetic data, three-stage training and inference optimization. On ALLEX humanoid tasks, RLDX-1 reaches 86.8% average success, while π0.5 and GR00T N1.6 reach about 40%.
Interpretation
RLDX-1 adds motion awareness, long-term memory and physical sensing as three separate functional capabilities to a robot policy, so one model can perform dexterous manipulation in dynamic, history-dependent and contact-rich tasks. Earlier vision-language-action models mainly inherit the general scene understanding and language-conditioned generalization of pretrained vision-language models, and are limited on tasks such as conveyor picking and the shell game. On the ALLEX humanoid benchmark, RLDX-1 reaches 86.8% average success while π0.5 and GR00T N1.6 reach about 40%; on conveyor pick-and-place, RLDX-1 reaches 100% at seen speeds and 75% at unseen speeds.
The model places each input modality in its own stream and lets those streams read from one another in attention layers, so each keeps its own representation while jointly generating actions. The Multi-Stream Action Transformer (MSAT) extends the two-stream structure of a multimodal diffusion transformer to action modeling, adding a physics stream when physical signals are available. The physics stream processes torque and tactile signals and additionally predicts future physical signals; on the Franka Research 3 platform, plug insertion succeeds at 33.3%, against 20.8% for π0.5 and 16.7% for GR00T N1.6.
Perspective
Evaluation covers simulation benchmarks including LIBERO, SIMPLER, RoboCasa Kitchen, GR-1 Tabletop and RoboCasa365, plus real tasks on the OpenArm and ALLEX humanoids and the Franka Research 3 single-arm platform; physical sensing is tested only on platforms carrying torque or tactile sensors.
Physical-signal data are much scarcer than visual observations and action labels, and tactile supervision on Franka Research 3 comes from a small set of in-house data; on unseen composite RoboCasa365 tasks the model reaches 5.6% success, still a low level.
