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

Taesung Kwon; Jangho Park; Sunwoo Park; Youngmin Kim; Seonghyun Jin; Youngjun Jun; Kyumin Choi; Jong Chul Ye

cs.ROSep 28, 2026

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

Authors: Taesung Kwon, Jangho Park, Sunwoo Park, Youngmin Kim, Seonghyun Jin, Youngjun Jun, Kyumin Choi, Jong Chul Ye

Organizations: KAIST · Sungkyunkwan University

Abstract

Modern robot policies predict a chunk of future actions from a single observation, execute only a prefix, and discard the rest before replanning. Choosing the length of this prefix, the execution horizon, poses a trade-off between reactivity and efficiency. A short horizon keeps the policy reactive to the environment, but requires frequent policy calls. Recent test-time methods adaptively select the horizon for each chunk, but they either read model internals, where the signal must be chosen for each architecture, or draw extra samples, which adds cost. We propose Action Upcycling, a training-free algorithm that reuses actions the policy would otherwise discard, without accessing model internals or drawing extra samples. We find that discarded actions stay close to their replanned versions as long as the action velocity remains smooth. Action Upcycling therefore extends the execution horizon up to the point where the velocity begins to fluctuate. Extensive experiments on simulated and real-world manipulation tasks show that Action Upcycling reduces policy calls by 1.2-1.7x with no loss in success rate, across multiple Vision-Language-Action Models (VLAs) and even a World Action Model (WAM). It applies to any chunked policy at negligible cost and is orthogonal to other policy acceleration methods such as few-step sampling and streaming action decoding, opening a new axis for policy acceleration.

Figures & tables

Figure 1: Discarded actions are worth keeping when motion is smooth. We compare the discarded actions of each chunk with their replanned versions on LIBERO for (a) SmolVLA, (b) GR00T N1.7, (c) $\pi_{0.5}$ , and (d) FastWAM. ( left ) They are strongly correlated. ( right ) Their RMSE grows with the velocity fluctuation of the discarded actions. Action Upcycling exploits this property to decide how many discarded actions to reuse.

Figure 2: Overview of Action Upcycling. Top : A chunked policy predicts

H

actions per call, executes only the first

h

, and discards the rest (the tail ). Action Upcycling additionally executes the trustworthy part of the tail, raising the mean execution length from

h

rh

and reducing policy calls accordingly. Bottom : The accumulated velocity fluctuation

c_{k}

stays small while the predicted motion is smooth and rises once it fluctuates. Action Upcycling executes tail actions while

c_{k}\leq\tau

, where

\tau

is determined from a pool

\mathcal{C}

of signals

c_{k}

, so that the mean execution length reaches

rh

Benchmark	Model	succ. $\uparrow$	calls / ep $\downarrow$	ms	s / ep $\downarrow$	succ. $\uparrow$	calls / ep $\downarrow$	ms	s / ep $\downarrow$
		Baseline Policy				+ Action Upcycling
LIBERO	$\pi_{0.5}$	96.9	32.4 (1 $\times$ )	136	4.41	97.9	21.9 (1.5 $\times$ )	137	3.00
	SmolVLA	82.6	19.3 (1 $\times$ )	94	1.82	82.8	14.7 (1.3 $\times$ )	94	1.39
	GR00T N1.7	96.1	22.6 (1 $\times$ )	115	2.59	96.6	15.0 (1.5 $\times$ )	115	1.72
	FastWAM	97.4	15.7 (1 $\times$ )	84	1.32	97.6	13.4 (1.2 $\times$ )	84	1.12
LIBERO-Plus	$\pi_{0.5}$	83.7	38.5 (1 $\times$ )	137	5.27	86.5	23.9 (1.6 $\times$ )	137	3.27

Table 1: Benefit of Action Upcycling. Results of four policies on three benchmarks. We report success rate (%), policy calls per episode, latency per policy call (ms), and inference time per episode (s). Action Upcycling consistently reduces policy calls by 1.2–1.7

\times

while matching or even improving the success rate. Better value in bold .

Method	succ.	calls / ep	ms	s / ep	succ.	calls / ep	ms	s / ep	succ.	calls / ep	ms	s / ep
	LIBERO				LIBERO-Plus				RoboTwin 2.0
Baseline	96.9	32.4 (1 $\times$ )	136	4.41	83.7	38.5 (1 $\times$ )	137	5.27	59.7	40.9 (1 $\times$ )	152	6.22
+ AAC (CVPR 2026)	97.3	23.5 (1.38 $\times$ )	802	18.9	84.5	27.1 (1.42 $\times$ )	802	21.7	53.1	46.1 (0.89 $\times$ )	930	42.9
+ AutoHorizon (ECCV 2026)	97.4	22.3 (1.45 $\times$ )	136	3.03	85.5	24.5 (1.57 $\times$ )	136	3.33	56.8	28.0 (1.46 $\times$ )	152	4.26
+ Action Upcycling	97.9	21.9 (1.48 $\times$ )	137	3.00	86.5	23.9 (1.61 $\times$ )	137	3.27	60.5	29.8 (1.37 $\times$ )	152	4.53

Table 2: Comparison with adaptive execution horizon methods. Results of

\pi_{0.5}

on three benchmarks. We report success rate (%), policy calls per episode, latency per policy call (ms), and inference time per episode (s). Best in bold , second best underlined .

Setting	succ. $\uparrow$	calls / ep $\downarrow$	ms	s / ep $\downarrow$	speed-up $\uparrow$
Baseline ( $N{=}10$ )	96.9	32.4 (1 $\times$ )	136	4.41	1.0 $\times$
+ Action Upcycling	97.9	21.9 (1.48 $\times$ )	137	3.00	1.5 $\times$
$N{=}5$	96.9	32.4 (1 $\times$ )	97	3.14	1.4 $\times$
+ Action Upcycling	97.7	21.9 (1.48 $\times$ )	98	2.15	2.1 $\times$
$N{=}2$	96.8	32.0 (1 $\times$ )	69	2.21	2.0 $\times$
+ Action Upcycling	97.8	21.8 (1.47 $\times$ )	71	1.55	2.8 $\times$

Table 3: Action Upcycling stacks with denoising-step reduction. Results of

\pi_{0.5}

with

N

denoising steps on LIBERO. We report success rate (%), policy calls per episode, latency per policy call (ms), inference time per episode (s), and speed-up over the

N{=}10

baseline. Better value at each

N

in bold .

Setting	succ. $\uparrow$	calls / ep $\downarrow$	ms	s / ep $\downarrow$	speed-up $\uparrow$
Baseline	96.90	32.4 (1 $\times$ )	136	4.41	1.0 $\times$
+ FlashVLA	98.65	31.4 (1.03 $\times$ )	32	1.01	4.4 $\times$
+ FlashVLA & Action Upcycling	98.70	21.5 (1.51 $\times$ )	31	0.66	6.7 $\times$

Table 4: Action Upcycling stacks with streaming action decoding. Results of

\pi_{0.5}

with FlashVLA on LIBERO. FlashVLA denoises a rolling buffer of chunks at different noise levels for faster decoding. We report success rate (%), policy calls per episode, latency per policy call (ms), inference time per episode (s), and speed-up over

\pi_{0.5}

. Best in bold .

	Baseline		Fixed-length		Adaptive (ours)
Model	succ. $\uparrow$	$\bar{h}$	succ. $\uparrow$	$\bar{h}$	succ. $\uparrow$	$\bar{h}$
$\pi_{0.5}$	96.90	5.0	97.05	7.0	97.90	7.3
SmolVLA	82.60	10.0	81.55	13.0	82.80	13.3
GR00T N1.7	96.05	8.0	95.40	12.0	96.55	11.9
FastWAM	97.35	10.0	96.45	12.0	97.55	11.9

Table 5: Adaptive vs. fixed-length execution. Results of four policies on LIBERO. Fixed-length execution uses a length close to the mean execution length

\bar{h}

of Action Upcycling. We report success rate (%) and

\bar{h}

. Better value in bold .

Figure 3: Distribution of execution lengths. The execution lengths selected by Action Upcycling spread over a range rather than concentrating on a single value.

	Success rate $\uparrow$		Calls / ep $\downarrow$		Time / ep (s) $\downarrow$
Task	Baseline	Action Upcycling	Baseline	Action Upcycling	Baseline	Action Upcycling
Soccer ball $\rightarrow$ plate (layout 1)	10/10	10/10	26.9	26.8	14.4	17.4
Soccer ball $\rightarrow$ plate (layout 2)	10/10	10/10	36.5	20.7	19.0	12.9
Basketball $\rightarrow$ basket	10/10	10/10	31.3	24.9	16.5	15.8
Baseball $\rightarrow$ basket	10/10	10/10	32.1	25.0	17.1	16.1
Drawer: black ball $\rightarrow$ drawer	8/10	10/10	73.8	49.2	37.3	30.9

Table 6: Quantitative real-world results with

\pi_{0.5}

. Each task is evaluated over 10 episodes. We report success rate, policy calls per episode, and time per episode (s). Better value in bold .

	Success rate $\uparrow$		Calls / ep $\downarrow$		Time / ep (s) $\downarrow$
Task	Baseline	Action Upcycling	Baseline	Action Upcycling	Baseline	Action Upcycling
Soccer ball $\rightarrow$ plate (layout 1)	10/10	10/10	56.8	50.3	17.8	17.7
Soccer ball $\rightarrow$ plate (layout 2)	8/10	9/10	102.2	81.7	31.5	28.8
Basketball $\rightarrow$ basket	8/10	8/10	95.6	69.9	29.6	24.3
Baseball $\rightarrow$ basket	8/10	8/10	107.9	82.3	33.1	28.9
Total / mean	34/40	35/40	90.6	71.0	28.0	24.9

Table 7: Quantitative real-world results with GR00T N1.6. Each task is evaluated over 10 episodes. We report success rate, policy calls per episode, and time per episode (s). Better value in bold .

Figure 4: Qualitative real-world results. A drawer task with $\pi_{0.5}$ (top) and a pick-and-place task with GR00T N1.6 (bottom), each shown with the baseline policy and with Action Upcycling. Red and green boxes mark failure and success, respectively. In both tasks, the baseline policy times out before completion, while the same policy with Action Upcycling succeeds. Videos are available on the project page.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Model	LIBERO	LIBERO-Plus	RoboTwin 2.0
$\pi_{0.5}$	5 / 10	5 / 10	10 / 32
SmolVLA	10 / 50	10 / 50	10 / 50
GR00T N1.7	8 / 16	8 / 16	—
FastWAM	10 / 32	10 / 32	24 / 32

Appendix

Table 8: Horizon settings. Execution horizon

h

and prediction horizon

H

of each model and benchmark.

Figure 5: Real-robot setup. Two YAM follower arms with wrist-mounted Intel RealSense D405 cameras, a top-mounted Intel RealSense D435, and two YAM leader arms used for teleoperated data collection. All experiments use only the right follower arm.

	All chunks		2 % of chunks
Model	$\tau$	$\bar{h}$	$\tau$	$\bar{h}$
$\pi_{0.5}$	0.15	7.3	0.15	7.3
SmolVLA	0.65	13.3	0.64	13.2
GR00T N1.7	0.22	11.9	0.22	11.9
FastWAM	0.09	11.9	0.09	11.9

Appendix

Table 9: A small pool suffices for selecting

\tau

. Selected

\tau

and resulting mean execution length

\bar{h}

on LIBERO, using pools built from all chunks and from a random 2 % of them.

Method	succ. $\uparrow$	calls / ep $\downarrow$	ms / call $\downarrow$	s / ep $\downarrow$
Baseline	96.9	32.4 (1 $\times$ )	136	4.41
+ Action Upcycling (offline pool $\mathcal{C}$ )	97.9	21.9 (1.48 $\times$ )	137	3.00
+ Action Upcycling (online pool $\mathcal{C}$ )	97.7	21.0 (1.54 $\times$ )	150	3.15

Appendix

Table 10: Offline vs. online pool construction. Results of

\pi_{0.5}

on LIBERO. Best in bold .

Figure 6: Effect of the upcycling ratio. Success rate (%) of each policy on LIBERO as

r

increases from the default setting (

r{=}1

Figure 7: Additional qualitative real-world results with $\pi_{0.5}$ . Four drawer tasks, each shown with the baseline policy (top) and with Action Upcycling (bottom). Red and green boxes mark failure and success, respectively. In two tasks, the baseline policy times out before completion, while the same policy with Action Upcycling succeeds. In the other two tasks, both succeed, but the policy with Action Upcycling completes the task faster.

Figure 8: Additional qualitative real-world results with GR00T N1.6. Four pick-and-place tasks, each shown with the baseline policy (top) and with Action Upcycling (bottom). Red and green boxes mark failure and success, respectively. In two tasks, the baseline policy times out before completion, while the same policy with Action Upcycling succeeds. In the other two tasks, both succeed, but the policy with Action Upcycling completes the task faster.

Explore similar work

Staircase Policy: Streaming Inference for World-Action Models with Large Action Chunks
Sep 29, 2026Guoheng Sun, Chen Chen, Jin Wang +2Efficient World-Action Model Action Chunks
University of Maryland, College Park · Independent Researcher · Oxford Robotics Institute, University of Oxford
PolicyTrim: Boosting Intrinsic Policy Efficiency of Vision-Language-Action Models
Jun 21, 2026Xianghui Wang, Feng Chen, Wenbo Zhang +4Inference Latency Robotic Manipulation
1Sichuan University · 2Adelaide University · 3Beijing Institute of Technology
PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking
May 30, 2026Junnan Nie, Jiayi Li, Chenghao Liu +5Action Chunks Robot Policies
Peking University · Peking University. · JD Explore Academy. +1

Sep 29, 2026, cs.ROJ/K move · Enter open · S save

Staircase Policy: Streaming Inference for World-Action Models with Large Action Chunks

Guoheng Sun, Chen Chen, Jin Wang, Ang Li+1

University of Maryland, College Park · Independent Researcher · Oxford Robotics Institute, University of Oxford

Efficient World-Action Model Action Chunks Action Generation Streaming Inference World Models Chunk

Open paper AskPDF arXiv

World-Action Models (WAMs) improve robotic manipulation by conditioning action generation on predicted future observations, but future prediction adds further inference overhead to already expensive iterative action generation. Action chunking can amortize this cost over multiple actions, yet performance degrades over long execution horizons because later actions remain conditioned on stale observations. We introduce STAIRCASE POLICY, a streaming inference and training framework that turns a flow-matching VLA into a JEPA-style WAM and partitions a large action chunk into sub-chunks at staggered denoising stages. Near-term actions are executed as soon as they become available, while later actions continue to be refined. At each sub-chunk boundary, the future latent is re-predicted from the latest observation and used to update all unexecuted actions, enabling long-horizon execution without repeated full policy inference. The resulting future-prediction error can further serve as a signal for adaptive chunking. S-WAM achieves 97.7% on LIBERO and 87.9% on LIBERO-Plus, and improves performance across multiple policy backbones and real-robot tasks. It reaches 292.7 executed actions per second,

3.62\times

the throughput of conventional execution at comparable accuracy, while reducing time-to-first-action from 123.6 to 73.3 ms. With additional inference optimizations, throughput further increases to 642.9 actions per second.

Abstract

Explore similar work

Staircase Policy: Streaming Inference for World-Action Models with Large Action Chunks

PolicyTrim: Boosting Intrinsic Policy Efficiency of Vision-Language-Action Models

PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking