Source note

World-Language-Action Model for Unified World Modeling, Language Reasoning, and Action Synthesis

World Language ActionRobot Foundation ModelWorld ModelVision Language ActionLong Horizon Manipulation

WLA proposes a world-language-action robot model that predicts textual subtasks, future visual states, and actions in one policy. The reported WLA-0 prototype reaches strong simulation and real-robot results with 2B active inference parameters and about 40 ms latency on an NVIDIA RTX 5090.

  • Existing world-action models mainly predict future images, which gives useful physical supervision but weak high-level language planning for long-horizon tasks.
  • Existing vision-language-action models can follow language but often lack direct future-state supervision for physical dynamics.
  • This matters because long-horizon robot manipulation needs semantic progress tracking, memory, and fast action generation under changing visual conditions.
  • WLA takes images, text instructions, and robot state as input, then predicts a textual subtask, a compact physical-dynamics representation, and an action chunk.
  • An autoregressive Transformer backbone generates the language subtask and meta-query outputs that encode the physical transition.
  • A World Expert trains those meta-query outputs to predict a future visual frame, while an Action Expert maps the same transition signal plus proprioception to executable actions.
  • The World Expert can be removed during normal inference, so action generation keeps the training benefit of world prediction without paying image-generation cost at test time.
  • A test-time scaling mode samples multiple action chunks, predicts the future frame for each one, scores those imagined states with a value model, and executes the highest-scoring action chunk.
  • WLA-0 has 3.4B total parameters, uses about 2B active parameters at inference, and reports about 40 ms inference latency on an NVIDIA RTX 5090.
  • On RoboTwin 2.0, WLA-0 reports 92.94% success on Clean and 90.02% on Randomized, with no embodied pretraining; Lingbot-VA reports 92.90% and 91.50%, and Fast-WAM reports 91.88% and 91.78%.
  • On LIBERO, WLA-0 reports 98.6% average success across Spatial, Object, Goal, and Long; test-time scaling with 6 candidates and horizon 2 raises the average to 98.9%.
  • Removing the world-modeling loss lowers RoboTwin Clean from 92.94% to 90.98% and LIBERO average from 98.6% to 97.9%, which supports the claim that future-state supervision improves action learning.
  • On RMBench, WLA-0 reports 56.5% average success, compared with 28.5% for Mem-0, 13.3% for Fast-WAM, 7.3% for X-VLA, and 5.5% for pi-0.5; removing language subtask loss drops WLA-0 to 17.3%.
  • For unseen RoboTwin 2.0 tasks with video-only data, the excerpt reports average gains over the seen-action baseline: same-embodiment video reaches 34.4 / 30.0 versus 13.0 / 11.6, and cross-embodiment video reaches 28.8 / 27.4.