Source note

Coarse-to-Control: Action-Token Planning for Vision-Language-Action Models

Vision Language ActionAction Token PlanningAction TokenizationRobot Foundation ModelsLong Horizon Manipulation

Coarse-to-Control adds an internal coarse action-token plan before a VLA emits executable action tokens, improving long-horizon robot manipulation in simulation and real robot tests.

  • Direct VLA policies map images, language, and robot state straight to motor actions, so long tasks can fail when early action errors compound.
  • Textual or visual intermediate reasoning can describe goals or subgoals, while wrist pose, motion direction, gripper timing, and waypoint structure remain under-specified.
  • The problem matters because multi-stage manipulation needs a policy to keep future task structure while producing precise low-level control.
  • The model first predicts coarse planning tokens from the current observation, instruction, and proprioceptive state.
  • It then predicts executable action tokens conditioned on those planning tokens; only executable tokens are decoded into robot actions at inference.
  • A joint residual-VQ tokenizer encodes both coarse long-horizon plans and short-horizon executable action chunks into one shared discrete action vocabulary.
  • The coarse plan is built by compressing a future action horizon into K chunks: each chunk stores net relative motion and the final gripper state.
  • Training uses teacher-forced autoregressive next-token prediction over the concatenated plan-token and execution-token sequence.
  • On LIBERO, Coarse-to-Control reports 97.9% overall success, above OpenVLA-OFT at 97.1%, π0.5 at 96.8%, and π0 at 94.2%; suite scores are 98.8 Spatial, 100.0 Object, 97.8 Goal, and 95.0 Long.
  • On SimplerEnv-WidowX, it reports 83.3% overall success, above UD-VLA at 62.5%, F1 at 59.4%, CogACT at 51.3%, and π0 at 40.1%; task scores are 100.0 Put Spoon, 95.8 Put Carrot, 79.2 Stack Block, and 58.3 Put Eggplant.
  • In real-world tests with 50 demonstrations per task and 20 rollouts per task, the plan-based policy reports 62.5% average success over four manipulation tasks and the best result on 3 of 4 tasks.
  • The LIBERO planning-horizon ablation improves overall success from 96.45% with no plan to 97.55% at H_p=40 and 97.90% at H_p=160.
  • The tokenizer ablation reports 95.40% overall for Faster-AR, 96.60% with separate planning and execution tokenizers, and 97.90% with the shared joint-mode tokenizer; Long-suite success rises from 88.60% to 91.60% to 95.00%.