DynamicManip: Enabling Dynamic Manipulation from a Single Static Demonstration

Haoran Liao^1,*, Pengyue Wang^1,*, Shuoyu Chen^1,*, Kehan Cheng¹
Xuhang Chen¹, Yuhao Lin¹, Mu Lin¹, Zhizhao Liang¹
Xiaoyi Fan⁴, Chengyi Xing², Dan Niu³, Yi-Lin Wei¹, Wei-Shi Zheng¹

¹ Sun Yat-sen University ² Stanford University ³ Southeast University ⁴ Jiangxing Intelligence
^* Equal contribution.

Paper BibTeX Code Coming Soon

DynamicManip teaser showing static-to-dynamic data augmentation and dynamic-aware policy execution. — DynamicManip augments diverse dynamic manipulation data from a single static real-world task demonstration, then trains a responsive dynamic-aware policy for generalized execution.

Abstract

Dynamic manipulation is a critical capability for robots operating in complex and dynamic environments, where robots must interact with objects that are moving or require rapid adjustments. However, learning models for dynamic manipulation tasks faces two major challenges: 1) the combinatorial complexity of dynamic scenarios leads to substantial data requirements, 2) rapid variations in dynamics require real-time and accurate policy execution. In this paper, we propose DynamicManip to address these challenges through an efficient data augmentation pipeline and a low-latency imitation policy. We first propose a static-to-dynamic augmentation pipeline that synthesizes diverse dynamic manipulation demonstrations from a single static demonstration. Second, we introduce a dynamic-aware adaptive policy that adaptively adjusts its inference frequency according to task dynamics, enabling responsive and effective dynamic manipulation. Third, we build a dynamic task manipulation benchmark, which includes diverse dynamic tasks with an automatic evaluation system for scalable and consistent assessment. Extensive experiments in both simulation and the real world demonstrate that DynamicManip not only provides significant improvements in data efficiency but also achieves superior performance in dynamic manipulation tasks, with higher success rates and better dynamic responsibility.

Method

Static Demonstration to Dynamic Execution

The framework couples data synthesis with adaptive inference: geometry reconstruction makes trajectory edits coherent, while stage labels produced by augmentation supervise a policy that reacts faster when dynamics demand it.

Static-to-dynamic data augmentation pipeline. — **Pipeline of our static-to-dynamic data augmentation.** From a single static demonstration, we localize the manipulation object and target and reconstruct their geometry. We then extract action-based keyframes, where labels such as start, grasp, interact, and recover only serve as intuitive annotations for segmenting the source motion into reusable pieces. Dynamic trajectory editing then organizes stages such as Dynamic Target Alignment to synthesize diverse trajectories, which are generated through motion planning and replaying local interaction segments.

Geometric Reconstruction

Object meshes are aligned to observed point clouds, while robot point clouds are rendered from the URDF and joint states for consistent scene editing.

Keyframe Extraction

A small set of contact and transition keyframes decomposes one static episode into reusable motion phases.

Trajectory Editing

Object motion, phase anchors, and transition goals are varied to synthesize dynamic episodes with physical and spatial consistency.

Adaptive Policy

Stage-aware supervision lets the policy adjust denoising and action execution schedules for low-latency closed-loop control.

Benchmark

DynamicManip Benchmark

The benchmark extends RoboTwin 2.0 with dynamic-aware data collection and automatic evaluation for moving-object manipulation tasks.

DynamicManip simulation benchmark task visualization. — Simulation tasks cover dynamic tapping, belt picking, object catching, mole whacking, goal blocking, and related dynamic interactions.

Real-world robot setup and objects used in experiments. — Real-world experiments use dual AgileX Piper robots with parallel grippers and a RealSense L515 depth camera.

Data Collection

Simulation. A dynamic-aware heuristic planner tracks moving targets, predicts future positions, and triggers grasps or contacts at task-specific moments.

Real world. We collect one static demonstration and apply static-to-dynamic data augmentation to synthesize dynamic training data.

Automatic Evaluation

Simulation. Task-specific state machines diagnose success, stage failures, trajectory residuals, and failure reasons after each rollout.

Real world. During evaluation, a human operator creates the dynamics by manually pushing or placing objects, and the rollout outcome is assessed on the physical setup.

Videos

Dynamic Manipulation Visualizations

Compressed videos are organized by real-world task, data setting, and rollout outcome.

Real-World Clips

Experiments

Data Efficiency and Low-Latency Control

DynamicManip is evaluated in simulation and real-world settings against models trained directly on human demonstrations and a DP3 baseline.

Simulation Data Efficiency

Success rate comparison in simulation: human demonstrations vs. synthetic data.
Setting	Dynamic Tapping	Belt Picking	Object Catching	Mole Whacking	Goal Blocking
1 demo	27	0	8	3	12
10 demos	15	5	54	21	46
50 demos	19	31	37	63	94
200 demos	55	85	50	83	87
1 demo to 50 augmented episodes	45	44	70	63	88
1 demo to 200 augmented episodes	52	82	68	80	87

Real-World Data Efficiency

Success rate comparison in real-world experiments.
Setting	Dynamic Tapping	Belt Picking	Boat Loading	Bottle Catching
1 demo	6.67	30.0	0.00	0.00
50 demos	68.9	70.3	3.33	73.3
1 demo to 50 augmented episodes	61.1	56.7	23.3	76.7
1 demo to 200 augmented episodes	87.8	73.3	66.7	83.3

Data Collection Time Cost

Time cost comparison for data collection and augmentation.
Method	Time Cost
Human Collection
Teleop. (50 dyn. demos, 2 persons)	~100 mins
Teleop. (200 dyn. demos, 2 persons)	~400 mins
Teleop. (1 static demo, 1 person)	~30 s
Automated Augmentation
Aug. (50 eps, human-free)	~7 mins
Aug. (200 eps, human-free)	~30 mins

Real-Time Performance

Latency and success rate comparison between DP3 baseline and DynamicManip.
Method	Tapping T.	Tapping Suc.	Belt T.	Belt Suc.	Catching T.	Catching Suc.	Whacking T.	Whacking Suc.	Blocking T.	Blocking Suc.
DP3 baseline	60.0	44.8	45.0	78	61.4	28	44.4	38	44.6	58
DynamicManip	56.0	32.5	32.9	80	36.0	26	30.6	44	33.9	88

Ablation Study

Contribution of mesh reconstruction and robot rendering.
Object Mesh Reconstruction	Robot Reconstruction	Average Success Rate
Enabled	Enabled	76.7%
Disabled	Enabled	10.0%
Disabled	Disabled	6.67%

BibTeX

@inproceedings{liao2026dynamicmanip,
  title     = {DynamicManip: Enabling Dynamic Manipulation from a Single Static Demonstration},
  author    = {Haoran Liao and Pengyue Wang and Shuoyu Chen and Kehan Cheng and Xuhang Chen and Yuhao Lin and Mu Lin and Zhizhao Liang and Xiaoyi Fan and Chengyi Xing and Dan Niu and Yi-Lin Wei and Wei-Shi Zheng},
  year      = {2026}
}