AdaPolicy: Action-Driven Anticipatory Robot Policies via Next-Chunk Prediction

Fan Yang; Yuting Su; Chen Zhao; Fugui Fan; Peiguang Jing

cs.ROOct 25, 2026

AdaPolicy: Action-Driven Anticipatory Robot Policies via Next-Chunk Prediction

Authors: Fan Yang, Yuting Su, Chen Zhao, Fugui Fan, Peiguang Jing

Organizations: Tianjin University · University of Bologna

Abstract

Endowing robot policies with anticipatory capabilities has proven effective in improving control performance. However, existing methods predominantly rely on explicit supervision derived from future observations, in either pixel or latent space, at the cost of additional architectural and computational complexity. Departing from explicit future supervision, we ask whether such foresight can be elicited directly from the action signal. To this end, we propose AdaPolicy, an Action-driven anticipatory policy framework that shapes future-aware representations via cascaded next-chunk prediction. Specifically, an Online Action Head jointly produces the current action chunk and an evolved latent state. Building on this state, a cascaded Unrolled Action Head forecasts the next chunk, and the resulting loss flows back through the evolved state to instill implicit foresight into the online policy. Importantly, this training-only unroll requires no future-observation encoding, keeping AdaPolicy lightweight to train with zero inference-time overhead. Experiments on MetaWorld, RoboTwin 2.0, LIBERO, and real-world deployments demonstrate consistent gains across diverse settings, including up to $14$ percentage points on millimeter-precision tasks such as earphone assembly and shoelace threading. More details and visualizations are available at https://teee111.github.io/AdaPolicy/.

Figures & tables

Figure 1: Comparison of future-supervision signals for robot policies. Existing methods supervise the predicted future state with explicit targets derived from future observations, including pixel reconstruction, latent alignment, and structured geometric targets (e.g., future object positions). In contrast, AdaPolicy uses the next action chunk as a control-aligned supervision signal, requiring no explicit future-observation targets.

Figure 2: Architectural overview. Left: The Context Adapter compresses pretrained features into compact conditioning tokens. Right: The Online Action Head jointly predicts the current action chunk and an evolved latent state, which conditions the Unrolled Action Head to predict the next chunk. During training, the next-chunk loss propagates through the evolved state to the online pathway; the cascaded step is omitted at inference time.

Figure 3: Visualization of real-world tasks. Representative keyframes of the sub-tasks are shown for the two challenging millimeter-precision tasks (top: earphone assembly, shoelace threading); the remaining four tasks are shown below.

Method	Earphone Assembly	Shoelace Threading	Stacking Blocks	Stacking Cups	Button Pressing	Lid Opening	Avg.
Baseline	74	36	82	100	70	100	77.0
AdaPolicy	86	50	94	100	88	100	86.3

Table 1: Success rates on real-world tasks. Each task is evaluated over 50 rollouts with randomized initial configurations.

Method	TinyVLA [ 74 ]	$\pi_{0}$ [ 75 ]	SmolVLA [ 76 ]	VPP [ 24 ]	RoMAN-Flow [ 77 ]	AdaPolicy
Avg.	31.6	50.8	68.2	68.2	72.8	81.2

Table 2: Success rates on MetaWorld in comparison with prior methods, averaged over all 50 tasks.

(a) 50 tasks, 50 clean + 500 randomized demos per task
Method	$\pi_{0}$	X-VLA	ABot-M0	Baseline	AdaPolicy
Avg. ( clean )	65.9	72.8	81.2	64.4	84.92
Avg. ( randomized )	58.4	72.8	80.4	64.5	84.04

Table 3: Average success rates on RoboTwin 2.0. (a) Comparison on 50 tasks, evaluated under both clean and randomized scenes (50 rollouts per task). (b) Ablation over backbones on 10 tasks, evaluated on clean scenes (50 rollouts per task).

Method	Goal	Object	Spatial	Long	Avg.
OpenVLA [ 1 ]	79.2	88.4	84.7	53.7	76.5
SmolVLA [ 76 ]	91.0	94.0	93.0	77.0	88.8
DreamVLA [ 78 ]	89.5	94.0	97.5	89.5	92.6
GR00T-N1 [ 79 ]	94.4	97.6	93.0	90.6	93.9
$\pi_{0}$ [ 75 ]	96.8	98.8	95.8	85.2	94.1
Baseline	93.0	96.4	96.0	83.0	92.1

Table 4: Success rates on the four LIBERO suites in comparison with prior methods.

Figure 4: Illustration of the unrolled-head variants. Left: frozen unrolled head with gradients flowing to the online pathway. Middle: trainable unrolled head with gradients flowing to the online pathway. Right: trainable unrolled head with stop-gradient on the online pathway.

w/o next-chunk prediction
Variant	$h$	$\gamma$	EMA $m$	Avg.
Baseline	8	–	–	67.8
Baseline (extended horizon)	16	–	–	74.3
Baseline (extended horizon)	24	–	–	60.4
Baseline + future-feature supervision	8	–	–	68.8
w/ next-chunk prediction

Table 5: Ablation on MetaWorld (average success rate over 50 tasks). Next-chunk prediction surpasses both the extended-horizon baseline and explicit future-feature reconstruction; gradient flow through the evolved state and the EMA parameterization of the unrolled head are both critical.

(a) Loss weight $\gamma$ on MetaWorld ( $m{=}0.98$ )
$\gamma$	0	0.5	0.6	0.7	0.8	0.9
Avg.	67.8 †	76.4	74.4	76.4	76.7	75.9
(b) EMA momentum $m$ on MetaWorld ( $\gamma{=}0.8$ )
$m$	0.95	0.96	0.97	0.98	0.99	0.995
Avg.	76.0	76.2	76.7	76.7	81.2	81.4
(c) EMA momentum $m$ on RoboTwin 2.0

Table 6: Hyperparameter sensitivity (average success rate). (a) Loss weight

\gamma

on MetaWorld with EMA

m{=}0.98

(

{}^{\dagger}\gamma{=}0

degenerates to the baseline without next-chunk prediction); (b) EMA momentum

m

on MetaWorld with

\gamma{=}0.8

; (c) EMA momentum

m

on RoboTwin 2.0.

	MetaWorld				RoboTwin 2.0
	1-step	2-step			1-step		2-step
	$\gamma{=}0.8$	$\gamma{=}0.8$ , $\gamma_{2}{=}0.4$	$\gamma{=}0.8$ , $\gamma_{2}{=}0.6$	$\gamma{=}0.8$ , $\gamma_{2}{=}0.8$	$\gamma{=}0.5$	$\gamma{=}0.7$	$\gamma{=}0.5$ , $\gamma_{2}{=}0.25$	$\gamma{=}0.7$ , $\gamma_{2}{=}0.3$
Avg.	81.2	75.6	77.4	75.1	72.4	67.0	66.2	61.0

Table 7: Effect of unroll depth on MetaWorld and RoboTwin 2.0 (average success rate).

\gamma

and

\gamma_{2}

denote the loss weights of the first and second unroll steps.

Noise Sampling Strategy	Condition on $(\tau_{0},\tau_{1})$	RoboTwin 2.0	MetaWorld
Baseline (w/o next-chunk)	–	63.8	67.8
Higher next-chunk noise	$1-\tau_{1}>1-\tau_{0}$ ( $\tau_{1}<\tau_{0}$ )	63.8	76.2
Lower next-chunk noise	$1-\tau_{1}<1-\tau_{0}$ ( $\tau_{1}>\tau_{0}$ )	69.2	79.0
Shared noise timestep	$\tau_{1}=\tau_{0}$	–	75.4
Independent sampling (AdaPolicy)	$\tau_{0}\perp\tau_{1}$ (default)	72.4	81.2

Table 8: Effect of relative flow-matching noise levels between the unrolled head (

1-\tau_{1}

) and the online head (

1-\tau_{0}

) on RoboTwin 2.0 and MetaWorld.

Temporal Offset $d$	Next-chunk Interval	Avg.
16	$[t+16,\ t+48)$	58.6
32 (default)	$[t+32,\ t+64)$	61.8
48	$[t+48,\ t+80)$	59.4

Table 9: Effect of the temporal offset

d

of the next-chunk prediction on RoboTwin 2.0. The action chunk size is

h=32

, and the default setting is

d=32

Model	$h$	Probed features	RoboTwin 2.0	MetaWorld
Baseline	8	condition state	16.38 $\pm$ 0.12	2.89 $\pm$ 0.09
Baseline	8	evolved state	15.76 $\pm$ 0.07	2.67 $\pm$ 0.05
Baseline + future-feature sup.	8	condition state	15.10 $\pm$ 0.19	2.78 $\pm$ 0.05
Baseline + future-feature sup.	8	evolved state	14.07 $\pm$ 0.23	2.55 $\pm$ 0.05
Baseline + extended horizon	16	condition state	15.28 $\pm$ 0.16	3.00 $\pm$ 0.07
Baseline + extended horizon	16	evolved state	14.06 $\pm$ 0.25	2.73 $\pm$ 0.09

Table 10: Linear probing of future object positions (average test-set error in pixels over future horizons,

\downarrow

, mean

\pm

std over 3 seeds).

	Cosine similarity		Attention mass
	$\cos(\hat{\mathbf{z}}_{t+h},\ \mathbf{c}_{t})$	$\cos(\hat{\mathbf{z}}_{t+h},\ \mathbf{c}_{t+h})$	cond $\to$ action
	(current)	(next)	(avg.)
Baseline	0.9977	0.6458	0.148
AdaPolicy	0.9650	0.9424	0.195

Table 11: Representation analysis on MetaWorld. Left: cosine similarity between the evolved state

\hat{\mathbf{z}}_{t+h}

and the condition states of the current and the next chunk. Right: attention mass from condition tokens to action tokens inside the online head, averaged over DiT layers and flow-matching denoising timesteps.

Method	reach@2cm	reach@3cm	reach@5cm
Baseline	6.0	8.2	17.8
AdaPolicy	12.8	15.5	29.5

Table 12: Latent-space planning with CEM on MetaWorld. We report the rate of reaching within

2/3/5

cm of the goal.

Figure 5: Per-step prediction error across successive chunks. Normalized L2 error on RoboTwin 2.0 for the current chunk (

A

), the trained unroll (

B

), and the second unroll (

C

), shown for the baseline (left) and AdaPolicy (right). Dashed lines mark the reference error of copying the previous ground-truth chunk.

Panel of Figure 6 — Figure 6: How does next-chunk supervision affect optimization? (a) Next-chunk supervision consistently lowers the current-chunk training loss; (b) the gradients of the two objectives stay positively correlated (moving average around $0.2$ ), indicating no systematic gradient conflict; (c) next-chunk prediction reduces the per-timestep action error throughout the chunk, with the largest margin at the final step.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Comparison of future-supervision architectures. Existing methods supervise a predicted future state with explicit targets derived from future observations, either by pixel reconstruction or by feature alignment. AdaPolicy instead derives supervision from the next action chunk, requiring no future-observation targets.

(a) Hidden size $512$
Variant	ms/step	Memory	Params	GFLOPs
AdaPolicy	1238	17.1 GB	93.1M	648.2
Chunk-32 baseline	1236	13.7 GB	93.1M	611.5
Chunk-64 baseline	1263	14.9 GB	93.1M	623.6
Future-feature sup.	2415	17.2 GB	111.7M	1163.8
(b) Hidden size $768$

Appendix

Table 13: Training cost (batch size 128) under two hidden sizes. AdaPolicy adds only modest computation over the baseline, whereas explicit future-feature supervision nearly doubles the per-sample FLOPs.

Figure 8: Real-world deployment with asynchronous inference. Inference of the action chunk overlaps with the execution of the current one, and B-spline trajectory refitting smooths the transition across chunk boundaries.

Figure 9: Examples of randomized initial configurations in real-world evaluation.

Figure 10: Visualization of the random initial configurations in the earphone assembly task.

Figure 11: Recovery from sub-optimal states on the earphone assembly task. Red arrows indicate the earbud tip, and green arrows mark the insertion slot of the charging case. Starting from a misaligned insertion attempt, the policy progressively re-aligns the earbud and successfully inserts it into the case.

Figure 12: Recovery from sub-optimal states on the shoelace threading task. Blue arrows denote the adjustment direction. Starting from a misaligned initial attempt, the policy progressively re-aligns the shoelace and successfully threads it through the eyelet.

Model	Probed features	$t{+}16$	$t{+}32$	$t{+}48$	Avg.
Baseline ( $h=32$ )	condition state	11.85 $\pm$ 0.13	19.00 $\pm$ 0.41	18.29 $\pm$ 0.22	16.38 $\pm$ 0.12
Baseline ( $h=32$ )	evolved state	11.60 $\pm$ 0.13	18.05 $\pm$ 0.05	17.62 $\pm$ 0.27	15.76 $\pm$ 0.07
Baseline ( $h=32$ ) + future-feature sup.	condition state	10.95 $\pm$ 0.18	16.93 $\pm$ 0.59	17.43 $\pm$ 0.17	15.10 $\pm$ 0.19
Baseline ( $h=32$ ) + future-feature sup.	evolved state	10.36 $\pm$ 0.10	15.75 $\pm$ 0.70	16.11 $\pm$ 0.24	14.07 $\pm$ 0.23
Baseline ( $h=64$ )	condition state	11.02 $\pm$ 0.17	17.26 $\pm$ 0.67	17.57 $\pm$ 0.08	15.28 $\pm$ 0.16
Baseline ( $h=64$ )	evolved state	10.41 $\pm$ 0.09	16.06 $\pm$ 0.82	15.70 $\pm$ 0.02	14.06 $\pm$ 0.25

Appendix

Table 14: Linear probing of future object displacement on the RoboTwin 2.0 stack-blocks-two task (test-set midpoint error in pixels,

\downarrow

, mean

\pm

std over 3 seeds).

Model	Probed features	$t{+}$ 4	$t{+}$ 8	$t{+}$ 12	$t{+}$ 16	Avg.
Baseline ( $h=8$ )	condition state	1.44 $\pm$ 0.03	2.54 $\pm$ 0.06	3.42 $\pm$ 0.10	4.16 $\pm$ 0.15	2.89 $\pm$ 0.09
Baseline ( $h=8$ )	evolved state	1.33 $\pm$ 0.02	2.32 $\pm$ 0.04	3.16 $\pm$ 0.06	3.87 $\pm$ 0.09	2.67 $\pm$ 0.05
Baseline ( $h=8$ ) + future-feature sup.	condition state	1.42 $\pm$ 0.02	2.48 $\pm$ 0.05	3.29 $\pm$ 0.06	3.93 $\pm$ 0.06	2.78 $\pm$ 0.05
Baseline ( $h=8$ ) + future-feature sup.	evolved state	1.33 $\pm$ 0.04	2.26 $\pm$ 0.06	2.99 $\pm$ 0.06	3.61 $\pm$ 0.05	2.55 $\pm$ 0.05
Baseline ( $h=16$ )	condition state	1.60 $\pm$ 0.04	2.72 $\pm$ 0.06	3.55 $\pm$ 0.10	4.12 $\pm$ 0.11	3.00 $\pm$ 0.07
Baseline ( $h=16$ )	evolved state	1.48 $\pm$ 0.05	2.48 $\pm$ 0.07	3.22 $\pm$ 0.09	3.73 $\pm$ 0.15	2.73 $\pm$ 0.09

Appendix

Table 15: Linear probing of future object displacement on 10 MetaWorld tasks (test-set midpoint error in pixels,

\downarrow

, mean

\pm

std over 3 seeds). Protocol follows Table 14 .

Setting	Embodiment	Action space	Camera views
RoboTwin 2.0	ALOHA-AgileX bimanual	14-D absolute joint + gripper (7-D per arm: 6 joints + 1 gripper)	one fixed third-person (head) camera
MetaWorld	Sawyer single arm	4-D end-effector displacement + gripper	two third-person cameras (corner, behindGripper)
LIBERO	Franka Panda single arm	7-D operational-space delta + gripper	wrist camera only
Real world	AgileX Piper single arm	7-D absolute joint + gripper (6 joints + 1 gripper)	two fixed third-person RealSense D435 cameras

Appendix

Table 16: Embodiment, action space, and camera setup across the four evaluation settings.

Category	Setting	MetaWorld & LIBERO	RoboTwin 2.0
Input	Image resolution	$256{\times}256$	$320{\times}240$
Architecture	Feature dimension	512
	Adapter query number	32	64
	Adapter layers	4
Optimization	Optimizer	AdamW
	Learning rate	$2{\times}10^{-4}$

Appendix

Table 17: Implementation details of AdaPolicy. Settings shared across benchmarks are shown as single entries.

Unrolled Head Variant	Next-Chunk Gradient Norm (Mean $\pm$ Std)	Variance
w/o EMA (detached online copy)	$0.3430\pm 0.1140$	$0.01300$
w/ EMA (AdaPolicy)	$\mathbf{0.2818\pm 0.0829}$	$\mathbf{0.00688}$

Appendix

Table 18: Comparison of next-chunk gradient statistics with and without EMA. Maintaining an EMA-updated unrolled head substantially reduces the variance of the gradients backpropagated from the next-chunk loss

\mathcal{L}_{\mathrm{next}}

	lever-pull	basketball	peg-insert-side	stick-push	pick-out-of-hole	sweep	shelf-place	Avg.
Baseline	54	34	14	10	2	4	4	17.4
AdaPolicy	64	24	16	26	16	12	4	23.1

Appendix

Table 19: CEM planning over frozen condition states (

c_{t}

) on seven MetaWorld tasks. A separate latent dynamics predictor is trained on top of the frozen

c_{t}

tokens to plan toward a goal

20

environment steps ahead.

	X-VLA		AdaPolicy			X-VLA		AdaPolicy			X-VLA		AdaPolicy
Task	Clean	Rand.	Clean	Rand.	Task	Clean	Rand.	Clean	Rand.	Task	Clean	Rand.	Clean	Rand.
Adjust Bottle	100	99	100	100	Open Microwave	79	71	82	58	Place Object Stand	86	88	84	94
Beat Block Hammer	92	88	94	80	Pick Diverse Bottles	58	36	72	84	Place Phone Stand	88	87	82	88
Blocks Ranking RGB	83	83	100	96	Pick Dual Bottles	47	36	92	92	Place Shoe	96	95	100	100
Blocks Ranking Size	67	74	62	74	Place A2B Left	48	49	58	76	Press Stapler	92	98	98	96
Click Alarmclock	99	99	92	88	Place A2B Right	36	36	56	54	Put Bottles Dustbin	74	77	90	92

Appendix

Table 20: Per-task success rates on 50 RoboTwin 2.0 tasks under clean and randomized settings.

Robot Policy Learning

Robot policy learning develops controllers that map sensory observations, task goals, and interaction history to actions for embodied robots.

See all 372 papers

FutureRTC: Real-Time Robot Execution with Anticipatory-Conditioned Action Chunking
Jul 27, 2026Hai Jiang, Yixian Zou, Binbin Liang +3Vision-Language-Action Models Action Chunking
School of Aeronautics and Astronautics, Sichuan University · University of Electronic Science and Technology of China · University of Alberta
Commit While Futures Agree: Consequence-Aware Adaptive Action Chunking for Robot Manipulation
Oct 6, 2026Yuyan Li, Yujia Wang, Yusong Huang +8Robotic Manipulation
The Hong Kong University of Science and Technology (Guangzhou) · ANU Intelligent · Victoria University of Wellington +2
Don't Throw Away the Tail: Action Upcycling for Policy Acceleration
Sep 28, 2026Taesung Kwon, Jangho Park, Sunwoo Park +5Robot Policy Learning Efficient VLA Model Inference
KAIST · Sungkyunkwan University

Jul 27, 2026J/K move · Enter open · S save

FutureRTC: Real-Time Robot Execution with Anticipatory-Conditioned Action Chunking

Hai Jiang, Yixian Zou, Binbin Liang, Boqian Liu+2

School of Aeronautics and Astronautics, Sichuan University · University of Electronic Science and Technology of China · University of Alberta

Vision-Language-Action Models Action Chunking VLMs for Robotics Robustness of VLA Models

Open paper AskPDF arXiv

Real-time deployment of Vision-Language-Action (VLA) policies necessitates asynchronous execution, wherein subsequent action chunks are computed concurrently with the execution of the current chunk, leading to prediction-execution misalignment and manifesting as inter-chunk discontinuities. Existing methods either superficially smooth chunk boundaries, require costly policy optimization, or exclusively forward-predict proprioceptive states yet neglect critical visual observations. In this paper, we propose FutureRTC, a plug-and-play adaptation framework that predicts execution-time observations and states for asynchronous VLA control without modifying the underlying policy. Specifically, FutureRTC features a state correction module to compensate for the discrepancy between rolled-forward and actual execution-time proprioceptive states and an observation prediction module that forecasts execution-time visual representations by leveraging robot motion as an explicit physical prior through motion-aware feature transport and reconstruction. Furthermore, we introduce a policy consistency loss to align the action chunks generated from predicted contexts with those produced under the expected execution-time inputs of the VLA policy. Extensive experiments across simulated and real-world environments demonstrate that FutureRTC achieves superior robustness to inference delays, resulting in smoother trajectories, faster execution, and consistently higher task success rates.

Abstract

Robot Policy Learning

Explore similar work

FutureRTC: Real-Time Robot Execution with Anticipatory-Conditioned Action Chunking

Commit While Futures Agree: Consequence-Aware Adaptive Action Chunking for Robot Manipulation

Don't Throw Away the Tail: Action Upcycling for Policy Acceleration