Revision, Not Restart: Revisable Visual Plans for Closed-Loop World-Action Models

Pengyiang Liu; Junbo Niu; Wenhao Zheng; Xinchen Chen; Canyu Li; Zhongyue Shi; Jiahao Xie; Si Liu

cs.ROSep 28, 2026

Revision, Not Restart: Revisable Visual Plans for Closed-Loop World-Action Models

Authors: Pengyiang Liu, Junbo Niu, Wenhao Zheng, Xinchen Chen, Canyu Li, Zhongyue Shi, Jiahao Xie, Si Liu

Organizations: Beihang University · Peking University

Abstract

World-action models use predicted visual futures to condition robot actions, yet execution feedback can invalidate parts of a prediction while leaving its task structure useful. We propose Revisable Temporal Planning (RTP), which maintains the visual future as a persistent action condition and revises it after feedback. Its central mechanism is a learned revision bridge: it resumes an intermediate state saved during visual generation and adapts its continuation to current observations. Visual and action supervision connect this revision to subsequent control. Time-aware history supplies observed evidence, and an adaptive policy selects retention, bridge revision, or fresh replanning from new noise before decoding the next action. On RoboMME and RMBench, RTP achieves task-averaged success rates of 48.6% and 84.8%, respectively. Matched comparisons support learned continuation; estimated checkpoint-source and action-prefix effects are positive but less precisely resolved. These results connect feedback-driven visual-plan revision to closed-loop task performance. Project Page: https://PLACEHOLDER.github.io/RTP/

Figures & tables

Figure 2 : The RTP feedback loop. The upper row illustrates a fresh plan root. At each feedback boundary, update factual history, compare prediction with feedback, and run one selected visual update. Save the accepted record and decode a new action from its next prefix and current facts.

Figure 3 : Prediction and execution. Base-model continuation under a joint-target hold at samples 52–59. Prediction (a) and execution observation (b) disagree at 60; the updated prediction for 76 (c) better matches the approach progress in the subsequent observation (d).

Method	Success (%) $\uparrow$
Vision–language–action models
$\pi_{0.5}$ ( Physical Intelligence et al., 2025 ; Dai et al., 2026 )	17.9
MME-VLA (TTT–Modul) ( Dai et al., 2026 )	22.0
MemER ( Sridhar et al., 2025 ; Dai et al., 2026 )	42.4
MME-VLA (FrameSamp–Modul) ( Dai et al., 2026 )	44.5
World–action models

Table 1: Task-averaged success on RoboMME.

Method	Success (%) $\uparrow$
Vision–language–action models
$\pi_{0.5}$ ( Physical Intelligence et al., 2025 ; Chen et al., 2026b )	10.4
X-VLA ( Zheng et al., 2025 ; Chen et al., 2026b )	9.8
Mem-0 ( Chen et al., 2026b )	42.0
World–action models
Fast-WAM ( Yuan et al., 2026 ; Yang et al., 2026 )	5.9

Table 2: Task-averaged success on RMBench.

Table 5

Action facts	Action prefix	Success (%)	Prefix effect (pp)	95% CI
Previous	Retained	41.3	1.8	[-0.9, 4.4]
Previous	Revised	43.0	1.8	[-0.9, 4.4]
Current	Retained	44.9	2.5	[0.0, 5.0]
Current	Revised	47.4	2.5	[0.0, 5.0]

Table 5: Action-input ablation under the Fixed bridge-10 policy. Prefix effects compare revised minus retained using unrounded counts; intervals are paired.

Figure 4 : Paired success differences on RoboMME. Dots and bars show differences and paired 95% intervals for the labeled policy contrasts, which are not additive. Sources: Tables 10 , 6 , and 5 .

Condition	I (%)	O (%)	N (%)	N–O (pp)	95% CI
Nominal	43.6	44.6	47.4	2.8	[-0.1, 5.5]
Four-sample hold	43.6	44.6	47.4	2.8	[-0.1, 5.5]
Eight-sample hold	43.0	44.2	47.1	2.9	[0.0, 5.6]

Table 6: Fixed-depth source comparisons under actuator holds, 800 keys/condition. I/O/N use independent-noise reconstruction, root-noise reconstruction, and saved checkpoints, respectively, with ten-interval continuation and fresh at exhaustion; difference and paired interval: N minus O.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Quantity	RoboMME	RMBench
Future / consumed groups	$4/1$	$4/1$
Native samples per block $J$	4	16
History budget $B_{H}$	60	72
Reset anchor quota	1	1
Recent nonanchor quota	12	16
Task-reference budget $B_{M}$	$0$ or $32$	0

Appendix

Table 7: History and model configuration. History-group budgets include all views; task references have a separate budget.

Task	Fresh	RTP	Rescues	Regressions	$\Delta$ [95% CI]
T01	13	13	–	–	0.00 [–]
T02	12	16	7	3	8.00 [-4.0, 20.0]
T03	11	14	4	1	6.00 [-2.0, 14.0]
T04	19	24	9	4	10.00 [-4.0, 24.0]
T05	16	18	5	3	4.00 [-6.0, 16.0]
T06	23	33	13	3	20.00 [6.0, 34.0]

Appendix

Table 8: Per-task RoboMME comparisons, 50 paired keys/task. Successes, rescues, and regressions are counts; differences and intervals are in percentage points. – denotes unreported paired statistics.

Task	Fresh	RTP	Rescues	Regressions	$\Delta$ [95% CI]
R01	61	63	–	–	2.00 [–]
R02	79	83	9	5	4.00 [-3.0, 11.0]
R03	77	81	6	2	4.00 [-1.0, 10.0]
R04	73	82	11	2	9.00 [2.0, 16.0]
R05	79	88	10	1	9.00 [3.0, 15.0]
R06	80	86	8	2	6.00 [0.0, 12.0]

Appendix

Table 9: Per-task RMBench comparisons, 100 paired keys/task. Rescues and regressions are counts; differences and intervals are percentage points. – denotes unreported paired statistics.

Contrast	Difference	95% interval
Foundation configuration	6.88	[4.13, 9.63]
Extra feedback fitting	3.12	[0.50, 5.63]
RTP vs. matched Fresh-20+C	5.125	–
Learned vs. zero, Fixed bridge-10	5.75	[3.00, 8.50]
RTP vs. Fixed bridge-10	1.250	–
RTP vs. binary retain/fresh	2.250	–

Appendix

Table 10: Paired policy and intervention contrasts, in percentage points. Overlapping contrasts are not additive module contributions. – denotes unreported intervals.

Policy	Visual steps	Success (%)	Call (s)	RTP diff. (pp)
Fresh-5	5.00	34.25	0.976	14.38
Fresh-5+C	5.00	36.50	0.991	12.13
Fresh-10	10.00	38.12	1.066	10.50
Fresh-10+C	10.00	41.75	1.081	6.88
Fresh-12+C	12.00	42.50	1.123	6.13
Fresh-20+C	20.00	43.50	1.275	5.13

Appendix

Table 11: Fresh integration schedules and RTP. Visual work includes structural refreshes; differences are RTP minus row.

Policy	Success (%)	Mean steps	Call (s)
Fixed bridge-5	45.62	8.49	1.057
Fixed bridge-10	47.38	12.33	1.115
Binary retain/fresh	46.38	11.10	1.113
RTP	48.63	7.78	1.051

Appendix

Table 12: Fixed-depth, binary retain/fresh, and adaptive updates. Each policy includes exhaustion refreshes.

Fitting	Nominal	Hold 4	Hold 8
Default RTP	69/4822 (1.43%)	91/4542 (2.00%)	94/4252 (2.21%)
Without age/length	181/4726 (3.83%)	176/4392 (4.01%)	219/4141 (5.29%)
Matched parent refit	74/4838 (1.53%)	91/4533 (2.01%)	93/4257 (2.18%)
Recurrent fitting	24/4944 (0.49%)	25/4666 (0.54%)	24/4376 (0.55%)

Appendix

Table 13: Recurrent discrepancy checks on 320 episodes/condition, with common frozen tolerances. Entries are exceedances / selected reuse (percent).

Policy	Success (%)	Mean steps	Difference [95% CI]
Default RTP	48.63	7.78	0.00 [0.00, 0.00]
Without age/length	48.00	7.80	-0.625 [–]
Matched parent refit	49.38	7.79	0.750 [–]
Recurrent fitting	49.50	7.75	0.875 [–]
Recurrent $-$ matched refit	–	–	0.125 [-2.50, 2.75]

Appendix

Table 14: Complete-policy recurrent-fitting controls on 800 reset keys. The first four differences are row minus default RTP; the final row is their direct recurrent-minus-matched contrast, not an additional policy. Differences and intervals are in percentage points; – denotes unreported intervals.

Boundary	Reached	Reused	Fresh	Exceedances	Exceedance (%)
1	2444	2168	276	27	1.25%
2	2070	1593	477	23	1.44%
3	1516	1061	455	19	1.79%
4	1016	0	1016	0	N/A
Total	7046	4822	2224	69	1.43%

Appendix

Table 15: Diagnostics by feedback boundary within each root. Exceedances count selected reuse with either distance above tolerance. N/A denotes the undefined zero-reuse ratio

0/0

Candidate	Contact MAE	Order (%)	Visual error
Retain	3.02	59.69	0.1251
Fresh	2.01	74.38	0.0893
Reconstruction-10	2.43	70.31	0.0955
Bridge-10	1.63	79.06	0.0691

Appendix

Table 16: Prediction against each candidate’s own action-conditioned continuation, 320 complete windows per branch; contact mean absolute error (MAE) is in native samples.

Condition	Zero (%)	Learned (%)	Difference (pp)
Nominal	41.62	47.38	5.75
Hold 4	41.62	47.38	5.75
Hold 8	40.75	47.12	6.38

Appendix

Table 17: Zero versus learned residual at fixed Fixed bridge-10. Source construction, depth, and action interface are matched; differences are learned minus zero.

Training loss	Success (%)	Visual error	Full–row (pp)
All terms	47.38	0.0840	0.00
No observed-visual loss	44.38	0.1206	3.00
No action loss	44.88	0.0805	2.50
No fresh regularizer	46.12	0.0961	1.25

Appendix

Table 18: Training-loss ablations at fixed Fixed bridge-10. Success and visual error have different targets; differences are full minus row.

Hold	Fresh	Fresh-20+C	Adaptive recon.	RTP	Difference (pp)
None	40.38	43.50	45.62	48.63	3.00
4	40.38	43.50	45.62	48.63	3.00
8	39.62	43.12	44.75	48.25	3.50

Appendix

Table 19: Complete-policy comparisons under actuator holds, 800 keys/condition. Differences are RTP minus Adaptive recon. Recon. denotes reconstruction from independent noise.

Condition / policy	Mean steps	Mean call (s)
4 / RTP	8.28	1.060
4 / Adaptive recon.	8.24	1.065
8 / RTP	8.88	1.071
8 / Adaptive recon.	8.95	1.078
4 / Fixed bridge-10	12.33	1.115
4 / Fixed recon.-10	12.32	1.117

Appendix

Table 20: Controller work under actuator holds. Noninitial calls use each policy’s own denominator. Recon. denotes reconstruction from independent noise.

Condition	Adaptive root recon. (%)	RTP (%)	Root recon. steps	RTP steps	Difference (pp)
Nominal	46.12	48.63	8.07	7.78	2.50
Hold 4	46.12	48.63	8.65	8.28	2.50
Hold 8	45.50	48.25	9.27	8.88	2.75

Appendix

Table 21: Original-root reconstruction and RTP. Differences are RTP minus Adaptive root recon.; step counts include all refreshes. Root recon. denotes reconstruction with the original root noise.

Delay $D$	Fresh-20+C	Fresh-10+C	Adaptive recon.	RTP	Difference (pp)
0	43.50	41.75	45.62	48.63	3.00
1	40.88	40.25	42.38	46.62	4.25
2	37.50	37.62	38.62	43.38	4.75

Appendix

Table 22: Activation-delay comparisons on 800 reset keys per delay. Differences are RTP minus Adaptive recon. Recon. denotes reconstruction from independent noise.

Policy	Success (%)	Mean steps	Matched diff. (pp)
Fresh-20+C	38.75	20.00	6.25
Adaptive recon.	41.25	8.94	3.75
RTP	45.00	9.03	0.00
Fixed recon.-10	39.50	12.31	3.75
Fixed bridge-10	43.25	12.32	0.00

Appendix

Table 23: Spatial-shift comparisons on 400 reset keys. Complete-policy differences use RTP; fixed-depth differences use Fixed bridge-10. Recon. denotes reconstruction from independent noise.

Setting	Fresh	Fresh-20+C	Adaptive recon.	RTP	RTP–Adaptive recon. (pp)
$H=6,J=4$	42.12	45.00	47.25	50.62	3.38
$H=4,J=8$	37.62	40.12	41.38	44.12	2.75

Appendix

Table 24: Horizon and sample-spacing comparisons on 800 reset keys per interface. Entries are success percentages. Recon. denotes reconstruction from independent noise.

Policy	Mean (s)	P50 (s)	P95 (s)
Fresh	1.248	1.248	1.304
Fixed bridge-5	1.057	1.007	1.281
Fixed bridge-10	1.115	1.088	1.268
RTP	1.051	1.010	1.294
Shared initialization	1.277	1.278	1.335

Appendix

Table 25: Controller-call durations. Policy rows contain noninitial calls; the last row contains the 800 shared fresh initializations. Quantiles use nearest rank.

Policy	Updates	Retain	Bridge-5	Bridge-10	Fresh	s/episode
Fresh	17403	0	0	0	17403	28.430
Fixed bridge-5	17882	0	13718	0	4164	24.902
Fixed bridge-10	17947	0	0	13769	4178	26.293
RTP	17922	6574	4407	2139	4802	24.817

Appendix

Table 26: Mode counts and episode cost on 800 episodes. Episode time includes initialization and all noninitial calls.

Explore similar work

Achieve What You Imagined: Learning to Align Actions with Visual Plans
Sep 27, 2026Yuheng Qiao, Ziran Wei, Xiaohan Wang +5Contact-Rich Dexterous Manipulation Action Expert
KTH · Beihang University · The Hong Kong University of Science and Technology (Guangzhou) +2
From World Models to World Action Models: A Concise Tutorial for Robotics
Jul 1, 2026Xiaoxiong Zhang, Xiong Zeng, Wei ZhangWorld Models Action Prediction
School of Automation and Intelligent Manufacturing, Southern University of Science and Technology, Shenzhen, China · LimX Dynamics, Shenzhen, China
World Action Planner: Generalizable Decision-Making with Action-Conditioned World Models
Jul 30, 2026Xiangcheng Zhang, Yilun DuEfficient World-Action Model World Model Planning
Harvard University

Sep 27, 2026, cs.ROJ/K move · Enter open · S save

Achieve What You Imagined: Learning to Align Actions with Visual Plans

Yuheng Qiao, Ziran Wei, Xiaohan Wang, Daqiang Guo+4

KTH · Beihang University · The Hong Kong University of Science and Technology (Guangzhou)+2

Contact-Rich Dexterous Manipulation Action Expert World Models Physical Plausibility

Open paper AskPDF arXiv

World-action models can jointly predict future visual observations and robot actions. However, discrepancies may exist between their visual predictions and the consequences implied by generated actions. We observe that WAMs can often generate visually plausible task-completion outcomes before producing action sequences that reliably achieve them. Consequently, we treat the WAM-generated visual prediction as a goal-conditioned visual proposal rather than a directly executable plan. We use a frozen action-conditioned world model to predict action-conditioned consequences and construct feedback based on consistency between the two future predictions and alignment with the terminal goal. Leveraging this feedback, we employ Flow Policy Optimization (FPO) to optimize the action head of the WAM. This framework avoids online robot interaction and additional training of task-specific reward models. Across four real-world UR5 manipulation tasks, our method increases the mean success rate from 43.4% to 75.1%, compared with 61.4% for

π_{0.5}

. These results show that cross-model prediction discrepancy can provide useful feedback for improving robot policies under the evaluated manipulation tasks. Website: https://imagine-to-achieve.github.io/

Abstract

Explore similar work

Achieve What You Imagined: Learning to Align Actions with Visual Plans

From World Models to World Action Models: A Concise Tutorial for Robotics

World Action Planner: Generalizable Decision-Making with Action-Conditioned World Models