Shared Execution-Clock Drifting Policy for Dynamic Precision Manipulation

Zhenchen Dong; Qingran Wu; Jinna Fu; Jiaming Wu; Fulin Chen; Hongyu Yu; Yide Liu

cs.ROSep 20, 2026

Shared Execution-Clock Drifting Policy for Dynamic Precision Manipulation

Authors: Zhenchen Dong, Qingran Wu, Jinna Fu, Jiaming Wu, Fulin Chen, Hongyu Yu, Yide Liu

Organizations: The Hong Kong Polytechnic University. · University of California San Diego. · Zhejiang University. · Universit´e Paris-Saclay. · Shanghai University of Engineering Science.

Abstract

Manipulation under time constraints requires both accurate actions and an execution rhythm that matches the evolving scene. This becomes critical when a robot must intercept moving objects or complete a sequence of adjustments before a deadline. Although one-step policies reduce generation cost, their directly predicted action sequences leave temporal allocation implicit. We propose Shared Execution-Clock Drifting (SECD), which makes execution rhythm an explicit part of one-step action generation. Conditioned on an observation and a latent sample, the policy jointly predicts a progress-indexed action curve and a shared monotone clock that maps fixed control times to locations on the curve. Demonstration-derived alignment anchors this decomposition, which is trained jointly through drifting on the decoded actions. The resulting policy retains a fixed-rate control interface and requires one network evaluation. We evaluate SECD across four real-robot tasks with inference on NVIDIA Thor. Across 300 trials, it achieves 77.00% task-averaged success and outperforms the evaluated one-step baselines on every task, including 91% success in cup retrieval from a 16 m/min conveyor and 54% in restoring and folding a crumpled shirt within 90 s. A fixed-clock variant reaches 79% on the same conveyor protocol. Complementary state-based RoboMimic experiments, including cross-seed ablations on Transport and Square, further support the joint design of the temporal representation and demonstration alignment. Project page: https://secd-anonymous-ewn.pages.dev/

Figures & tables

Fig. 1: Shared Execution-Clock Drifting. A joint generator predicts a progress curve and a monotone clock, which decode to fixed-rate actions in one network evaluation. Demonstration alignment and decoded-action drifting jointly train both outputs. Curves and clock profiles are schematic.

Fig. 2: Real-robot hardware setup. A Franka Research 3 arm performs conveyor retrieval (left). The Flexiv Rizon 4s setup (right) uses one arm for cup placement and both arms for shirt manipulation. The insets show the gripper and wrist camera, example RGB observations, and the NVIDIA Jetson AGX Thor computer on which all compared policies run inference.

Fig. 3: Cup Place: placing a cup on a coaster. A single arm grasps and transports a stationary cup to a coaster whose position varies across trials.

Fig. 4: Shirt folding from flat and crumpled states. The upper row shows recovery from a crumpled state through flipping, collar grasping and lifting, lowering, and flattening. The lower row shows right- and left-sleeve folding, a half-fold, a $90^{\circ}$ rotation, and a second half-fold. Fold (Crumpled) requires the full sequence within $90$ s; Fold (Flat) starts from the flattened state.

Fig. 5: Conveyor Cup: retrieving a moving cup. The robot approaches and grasps a cup on a conveyor moving at $16$ m/min before withdrawing and releasing it. The timestamps indicate elapsed time in the illustrated execution sequence.

Method	NFE	Lat. (ms)	Cup Place	Fold (Flat)	Fold (Crumpled)	Conveyor Cup	Avg.
DP [ 1 ]	16	252	98.0	74.0	30.0	0.0	50.50
OneDP [ 4 ]	1	26	71.0	38.0	22.0	31.0	40.50
Native Drifting [ 7 ]	1	24	83.0	48.0	26.0	56.0	53.25
SECD (ours)	1	25	95.0	68.0	54.0	91.0	77.00

TABLE I: Real-robot success (%) and mean policy inference time on NVIDIA Thor. Each policy is evaluated on 100 Cup Place, 50 Fold (Flat), 50 Fold (Crumpled), and 100 Conveyor Cup trials. Avg. is the unweighted mean across tasks. Fold (Crumpled) requires flattening and full folding within

90

s; Conveyor Cup uses a belt speed of

16

m/min. Bold marks the highest success in each column.

Fig. 6: Cup grasping and placement failures. The examples show inaccurate placement (left), unstable grasping (center), and inaccurate grasping (right).

Fig. 7: Failures during shirt manipulation. The examples illustrate inaccurate grasping, collision, and overtravel that disrupt the manipulation sequence.

Fig. 8: Missed conveyor grasp with DP (16 NFE). The cup continues moving during the approach, and the gripper fails to intercept it.

Method	NFE	Lift	Can	Square	Transport	Avg.
DP [ 1 ] †	100	1.00 (1.00)	1.00 (0.98)	0.96 (0.92)	0.73 (0.62)	0.92 (0.88)
Flow [ 22 ] †	9	1.00 (1.00)	1.00 (0.98)	0.95 (0.90)	0.74 (0.59)	0.92 (0.87)
MIP [ 23 ] †	2	1.00 (1.00)	1.00 (0.98)	0.96 (0.88)	0.71 (0.58)	0.92 (0.86)
CP [ 3 ] †	1	1.00 (1.00)	0.99 (0.96)	0.88 (0.82)	0.32 (0.24)	0.80 (0.76)
MP1 [ 6 ] †	1	1.00 (1.00)	1.00 (0.97)	0.96 (0.88)	0.52 (0.39)	0.87 (0.81)
Native Drifting [ 7 ] †	1	1.00 (1.00)	0.99 (0.96)	0.97 (0.90)	0.68 (0.58)	0.91 (0.86)

TABLE II: State-based RoboMimic PH success: best (final-five-checkpoint average); Avg.: four-task mean.

\dagger

: IDP Table 2 [ 13 ] , averaging CNN/Transformer backbones, three seeds, and 50 initial conditions; NFE follows the source. Native Drifting denotes its “Naive Drifting.” SECD: local U-Net, 100 rollouts/checkpoint.

	Transport		Square
Variant	Best	Final	Best	Final
A: Direct drifting	0.68	0.58	0.97	0.90
B: Uniform clock + MSE	0.74	0.60	0.96	0.92
C: SECD w/o alignment	0.70	0.56	0.97	0.88
D: Complete SECD	0.78	0.68	0.97	0.90

TABLE III: Transport and Square ablations: success rates aggregated across training seeds. Best and Final denote best-checkpoint success and the final-five-checkpoint average, respectively.

Explore similar work

PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking
May 30, 2026Junnan Nie, Jiayi Li, Chenghao Liu +5Robotic Manipulation Action Chunking
Peking University · Peking University. · JD Explore Academy. +1
Chronos Policy: Learning Robot Policies on Physical Time
Oct 9, 2026Changhao Wang, Jaden Clark, Shuran SongRobot Policy Learning Robot Learning
Stanford University
Dynamic Execution Horizon Prediction for Chunk-based Robot Policies
Jun 9, 2026Yuchi Zhao, Miroslav Bogdanovic, Arjun Sohal +5Robot Policy Learning Long-Horizon Robotic Manipulation
Department of Computer Science, University of Toronto, 40 St George St., Toronto, ON M5S 2E4, Canada · Vector Institute for Artificial Intelligence, W1140-108 College St., Schwartz Reisman Innovation Campus, Toronto, ON M5G 0C6, Canada · Acceleration Consortium, 700 University Ave., Toronto, ON M7A 2S4, Canada +7

May 30, 2026, cs.ROJ/K move · Enter open · S save

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

Junnan Nie, Jiayi Li, Chenghao Liu, Junyi Lao+4

Peking University · Peking University. · JD Explore Academy.+1

Robotic Manipulation Action Chunking

Open paper AskPDF arXiv

Recent vision-language-action and diffusion-based robot policies often use action chunking, where each policy query predicts a sequence of future actions and the robot executes an open-loop prefix before re-querying. While this interface improves local motion continuity, deployment still requires choosing the execution horizon: how much of each predicted chunk should be executed before acquiring a new observation. However, our experiments show that success is strongly task-dependent and non-monotonic with respect to the execution horizon, making a single constant horizon an unreliable deployment rule. We propose PACE (Phase-Aware Chunk Execution), a training-free test-time execution method that selects the execution horizon online from the predicted chunk itself. PACE exploits the phase-dependent kinematic structure of manipulation trajectories by identifying low-speed transition points in the predicted speed profile and using them as candidate replanning boundaries. Because PACE uses only the predicted action chunk, it is plug-and-play and requires no retraining or access to policy internals. We validate PACE through large-scale evaluations in both simulation and real-robot settings. On 50 RoboTwin2.0 tasks, PACE raises the average success rate from 57.8% to 64.2%. In real-robot experiments on bimanual ALOHA and single-arm Franka platforms, PACE improves the average task score from 60.7 to 77.7 and the average success rate from 50.7% to 70.4%. Ablations and rollout-level analyses show that PACE adapts execution horizons across manipulation phases, shortening near transitions while preserving longer execution during coherent motion.

Abstract

Explore similar work

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

Chronos Policy: Learning Robot Policies on Physical Time

Dynamic Execution Horizon Prediction for Chunk-based Robot Policies