Source-linked AI summary
PhasePlan: Ordered Future-Phase Planning for Robot Brain Models
Xiaoyu Yang, Yafei Zhang, Wensheng Li, Qing Zhan, Nan Wu
TL;DR
Action chunks can span multiple task phases, making transition timing important, while phase counts alone do not preserve phase order. PhasePlan predicts future phases position by position and conditions corresponding actions on those plans. In conveyor-belt experiments, it reduces offline joint-action error by 22.5% relative to the original π0.5 policy.
Problem
Action chunks may cross task-phase boundaries, and sequences with the same phase counts can require different actions at each future position.
Source: §1 · p. 1Method
Results
Takeaways & Limitations
Conveyor-belt experiments show the design's effectiveness in the π0.5 implementation, including gains on action chunks that cross phase boundaries.
Source: §6 · p. 7
Abstract
from arXiv · showhide
Robot brain models integrate vision, language, and robot state to generate actions for complex manipulation tasks. Most predict fixed-length action chunks that may span multiple task phases. This can obscure phase transitions and favor frequent action patterns, compromising action timing in dynamic environments. We propose \method, an ordered future-phase planning method for robot brain models. From current multimodal observations, it predicts the task phase at each future action position. The resulting planning representations condition the corresponding actions, preserving temporal alignment between task progress and action generation. Training first learns the planner, then freezes it during action-model adaptation to maintain stable phase representations. We instantiate \method on pretrained $\pi_{0.5}$ and AcrossWAM1.0 robot brain models. Detailed quantitative evaluation uses the $\pi_{0.5}$ implementation. On conveyor-belt manipulation, \method reduces offline joint-action error by approximately 22.5\% relative to the original $\pi_{0.5}$ model. It also improves phase-transition modeling and cross-phase action prediction. These results demonstrate the value of ordered future-phase planning for continuous action generation.
1. Introduction
Action chunks can span task phases, making transition timing important beyond average action error; PhasePlan predicts ordered future phases and aligns each plan feature with its corresponding action. In conveyor-belt experiments, it reduces offline joint-action error relative to both the original policy and a future-skill-histogram baseline.
1. Introduction
- Action chunks can cross phase boundaries, so policies with similar average action errors may differ in transition timing and closed-loop behavior.Waiting and transport take longer than release and phase transitions, while action optimization favors frequent motion patterns.Source: §1 · p. 1
- Position-aligned phase representations distinguish sequences with identical phase counts but different phase order, such as Wait, Wait, Act, Act versus Act, Act, Wait, Wait.Source: §1 · p. 1
- PhasePlan predicts a phase distribution for each of H = 16 future positions, and each position's planning feature conditions its corresponding future action.Its inputs are head-view and wrist-view RGB, proprioceptive state, and a language instruction.Source: §1 · p. 1
- Training first learns the ordered future-phase planner, then freezes it while adapting the action expert and plan-conditioning path.The stated aim is to let action adaptation use learned phase-transition structure.Source: §1 · p. 1
- The independent-query planner achieves about 92.5% per-position future-phase accuracy and about 89% phase-transition-type accuracy.Source: §1 · pp. 1–2
- Offline joint-action error falls 22.5% relative to the original π0.5 and 12.7% relative to the future-skill-histogram baseline.The result uses a frozen planner with per-position conditioning.Source: §1 · p. 2
- The contributions combine position-corresponding phase predictions, matching plan conditioning for each future action token, and two-stage training with a frozen planner.The paper states that this maintains phase-boundary quality while improving action prediction.Source: §1 · p. 2
2. Related Work
Related work covers robot-brain and vision-language-action policies, temporal progress representations, and methods connecting phase information with future actions. PhasePlan distinguishes its approach by predicting phase order and transition positions at each future action position.
2.1. Robot Brain Models and VLA Policies
- Robot-brain and VLA advances provide a foundation for executable policies and downstream adaptation, which PhasePlan builds on [6].Source: §2.1 · p. 2
2.2. Temporal Planning for Robot Control
- Temporal-control approaches use language-action descriptions, behavior and phase representations, progress estimates, and progress-related belief states [1] [5] [16] [15].Source: §2.2 · p. 2
- Dense LSS, ForeTime-VLA, and GuidedVLA connect phase or future-skill information with actions, while PhasePlan predicts phase order and transition positions per action position [9, 10, 7].GuidedVLA predicts a future skill distribution describing phase proportions within an action chunk.Source: §2.2 · p. 2
3. Method
PhasePlan predicts an ordered phase plan from multimodal observations and aligns each future-position representation with its corresponding action. Its planner is trained first, then frozen while the action expert and conditioning path are adapted.
3.1. Overview and Problem Formulation
- Conveyor experiments use H = 16 positionwise phase labels, constructed from object annotations, gripper state, and task-specific temporal rules, to supervise an internal ordered future plan.At deployment, the policy generates phase distributions from its standard observations.Source: §3.1 · pp. 2–3
3.2. Ordered Future-Phase Planner
- Each of H learnable queries reads shared observation context while retaining its future-position identity, predicting a phase distribution for the corresponding action position.The per-position classification objective preserves phase categories and transition locations within the action chunk.Source: §3.2 · p. 3
3.3. Plan-Conditioned Action Generation
- At each future position, PhasePlan fuses a contextual planning representation with a soft phase embedding and uses the resulting feature to modulate the corresponding action token.Each feature affects only its matching action token, maintaining plan–action position alignment.Source: §3.3 · p. 3
3.4. Training and Inference
- Training first optimizes the phase objective with the action expert fixed, then freezes the planner while adapting the action expert and per-position modulation path.The frozen planner continues supplying planning representations and phase distributions during action adaptation.Source: §3.4 · pp. 3–4
4. Experiments
The experiments evaluate PhasePlan on a teleoperated conveyor task with position-aligned supervision for 16-step action windows. Comparisons use a shared π0.5 backbone and training setup to assess ordered planning and its effect on action generation.
4.1. Task and Experimental Setup
- Each cycle waits for an actionable red package, grasps and transports it to a blue basket, releases it, and returns the arm home.The task is represented by four phases, and each policy action step contains seven right-arm joint targets and a binary hand command.Source: §4.1 · p. 4
- The conveyor dataset contains 325 trajectory episodes and 40,129 frames, with phase labels aligned position by position to each valid 16-step action window.The split is fixed by episode; preprocessing removes stationary Wait-phase jitter and re-anchors the Wait-to-Act transition, with the same processed action targets used across methods.Source: §4.1 · p. 4
4.2. Baselines and Metrics
- PhasePlan is compared with original π0.5, an aggregated future-skill-histogram baseline, and Ours-P, which trains ordered phase prediction alongside the original action path.Methods share the π0.5 initialization, demonstrations, episode split, optimization budget, and action representation; ablations vary planner queries and inputs, action conditioning, planner updates, and oracle phase labels.Source: §4.2 · p. 4
5. Results
PhasePlan improves offline action prediction on the conveyor task, while its planner predicts future phases accurately and supports transition modeling. Simulation, physical-robot demonstrations, and ablations further assess execution and design choices.
5.1. Future-Phase Planning and Action Prediction
- The independent-query planner achieves about 92.5% per-position future-phase accuracy and about 89% phase-transition-type accuracy.Full vision-language-state input exceeds 91% macro-averaged F1, compared with about 43% for state and language alone and about 36% with shuffled RGB; independent queries perform slightly better than interactive queries on the main phase metrics.Source: §5.1 · p. 5
- Joint-action MAE is 0.034605 rad, 22.5% below original π0.5 and 12.7% below the future-skill-histogram baseline.On cross-phase windows, MAE is 0.041318 rad and hand-command accuracy is 93.162%; PhasePlan also has the lowest joint-action error on both the full validation set and cross-phase chunks.Source: §5.1 · p. 5
5.2. Simulation and Real-Robot Validation
- In MuJoCo, executing predicted actions with recorded visual inputs produces continuous arm-pose and hand-configuration changes.The setup uses recorded validation images and continuously updated simulated robot states, providing a controlled view of actions under simulated kinematics and physical constraints.Source: §5.2 · pp. 5–6
- A real conveyor-robot sequence shows coherent approach, grasp, transport, and placement under real visual feedback and robot dynamics.The sequence is qualitative and complements the offline metrics and MuJoCo results.Source: §5.2 · p. 6
- Qualitative AcrossWAM1.0 executions show progression from approach and grasp through transport and release from side and frontal viewpoints.The displayed sequences involve sorting a pink package on a green conveyor into a blue collection basket.Source: §5.2 · p. 6
5.3. Ablation Studies
- Full contextual plan features yield lower action error than phase-ID-only conditioning, supporting the use of both contextual representations and phase distributions.The combined signals form the per-position conditioning feature.Source: §5.3 · p. 6
- Freezing the planner improves phase-transition-type accuracy and transition timing while yielding slightly lower joint-action error; the future-phase oracle improves action prediction further.The remaining oracle gap is largest for Release-related hand predictions, highlighting accurate transition prediction for short manipulation phases.Source: §5.3 · pp. 6–7
- Inverse-phase-frequency weighting raises hand-command accuracy at future Release positions from about 82.6% to 84.3%, but increases overall joint-action MAE by about 1.3%.Cross-phase action error also increases slightly, so the final model uses the uniformly weighted objective for a better overall trade-off.Source: §5.3 · p. 7
6. Conclusion
PhasePlan represents task progress with ordered, position-aligned future-phase predictions that condition corresponding actions, while freezing the planner during action adaptation. Conveyor-belt experiments report lower offline joint-action error and coherent execution in simulated and physical settings.
6. Conclusion
- Each future action position receives a phase distribution and contextual planning representation that condition its corresponding action, preserving phase order and transition locations.The two-stage strategy first learns future-phase representations, then freezes the planner during action adaptation to maintain stable temporal semantics.Source: §6 · p. 7
- Offline joint-action error falls 22.5% relative to the original policy and 12.7% relative to the future-skill-histogram baseline in the π0.5 implementation.The conveyor-belt experiments also report gains on action chunks crossing phase boundaries.Source: §6 · p. 7
- The planner achieves about 92.5% per-position phase accuracy and 89% phase-transition-type accuracy.Source: §6 · p. 7
- Recorded-visual-input MuJoCo execution produces continuous motion as simulated states evolve, and physical deployment shows coherent progression from grasping to placement.AcrossWAM1.0 deployment also shows progression from approach and grasp to transport and release.Source: §6 · p. 7