PhasePlan: Ordered Future-Phase Planning for Robot Brain Models
This preprint introduces PhasePlan, an ordered future-phase planning method for robot brain models. On a real-robot conveyor sorting task with a pretrained π0.5 backbone, PhasePlan reduces offline joint-action error by 22.5% relative to the original policy and reaches 92.5% per-position phase accuracy. By preserving where phase transitions occur within each 16-step action chunk, it may improve action timing in dynamic manipulation tasks.
Can per-position future task-phase predictions, used to condition each corresponding future action token, improve action timing and accuracy in pretrained vision-language-action policies for manipulation?
Existing VLA policies often predict fixed-length action chunks that can span several task phases, such as wait, act, release, and return-home. These models learn average action patterns and can miss rare but critical phase transitions, so two policies with similar average action errors may behave differently in closed-loop control, especially on conveyor manipulation where waiting and transport dominate and release is brief.
Prior temporal-planning methods either adapt pretrained VLA action experts directly or use summarized progress signals such as intermediate language actions (RT-H), latent behavior or phase representations (BehaviorVLA, ProgressVLA, S2-VLA), and future-skill histograms (GuidedVLA). These approaches often aggregate phase information over an interval or inject it without per-position alignment, so they can lose the temporal order and exact position of phase transitions within an action chunk.
PhasePlan adds H learnable phase queries to a pretrained VLA. Each query attends to multimodal observation tokens (head-view RGB, wrist-view RGB, proprioception, and language) and predicts a phase distribution for one of 16 future action positions. The planner representation and soft phase embedding are fused and used to modulate the corresponding action-expert token through a zero-initialized FiLM-style path. Training is two-stage: first the planner is trained with position-wise phase classification while the action expert is held fixed, then the planner is frozen and the action expert plus modulation path are fine-tuned with the flow-matching objective. It is evaluated quantitatively on the π0.5 backbone and qualitatively on AcrossWAM1.0.
On conveyor validation (3,584 windows; 1,613 cross-phase; averaged over three seeds), PhasePlan overall joint MAE is 0.034605 rad versus 0.044653 for original π0.5 and 0.039619 for the future-skill histogram baseline, i.e., 22.5% and 12.7% relative reductions. Cross-phase MAE is 0.041318 rad and hand-command accuracy is 93.162%, compared with 90.394% for original π0.5. The planner achieves 92.5% per-position phase accuracy and 89% phase-transition-type accuracy. Ablations: removing or shuffling RGB reduces macro F1 to 43% and 36% from over 91%; box-level visual supervision raises attention inside annotated regions from 8.5% to 81.2%; inverse phase weighting improves release-hand accuracy from 82.6% to 84.3% but worsens overall joint MAE by 1.3%. MuJoCo and real-robot runs are qualitative only; no real-robot success-rate numbers are reported.
The paper does not include a separate limitations section. Evident limitations: quantitative results are limited to one conveyor task and one π0.5 backbone, with only qualitative deployment on AcrossWAM1.0; real-robot and MuJoCo evaluations report no success rates or closed-loop statistics; phase labels rely on object annotations and task-specific temporal rules; the dataset is 325 episodes; and no code or model release is stated.
Warehouse and logistics robotics teams that deploy vision-language-action policies for pick-and-place, conveyor sorting, or machine tending could use this as a fine-tuning module to improve transition timing without changing inference inputs. Pretrained VLA providers or industrial robot OEMs could integrate PhasePlan into model adaptation pipelines. Near-term use is likely research to 1–3 years if validated across more tasks, robots, and real-robot success metrics, and if phase-label extraction can be automated robustly. Broader adoption requires evidence that the 22.5% offline MAE gain translates to fewer missed grasps or releases in production settings.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.