CLAP: Closed-Loop Alignment with Pressure for Precise Suction Manipulation

Yixian Zou; Chongyang Xu; Yuling Xin; Ziliang Feng; Fanman Meng; Shuaicheng Liu

cs.ROSep 26, 2026

CLAP: Closed-Loop Alignment with Pressure for Precise Suction Manipulation

Authors: Yixian Zou, Chongyang Xu, Yuling Xin, Ziliang Feng, Fanman Meng, Shuaicheng Liu

Organizations: University of Electronic Science and Technology of China. · School of Aeronautics and Astronautics, Sichuan University.

Abstract

Stacking and palletising demand precise placement: error left in one layer is inherited by the next, and a flat pad offers no feature to funnel a wrong pose into the right one. Top-down suction suits such dense arrangements, and suction has already been brought into vision-language-action (VLA) policies. What that work does not report, however, is a policy conditioned on a measured vacuum signal, or one that uses it to abandon an action already under way. Vision does not settle the question here, because at the moment it matters the cup and the face it holds occlude each other. We present CLAP, which makes the attachment state observable through a pressure module tapped into the vacuum line. The decoded reading replaces the suction command in the policy's proprioception, is fused with the visual features, and terminates the open-loop execution window so that the policy re-infers from a fresh observation. For data, we record goal-state disassembly on the physical robot and reverse the joint-state sequence offline, without a simulation replay. Targeted phase demonstrations, 8.3% of the training frames, cover the suction transitions and the configurations an interrupted grasp leaves behind. On a real Unitree Z1, one multi-task checkpoint reaches 96.67% average success in both colour settings, 16.67 and 10.00 points above the strongest baseline, its monochromatic four-block successes averaging 15.92 mm of error. Four ablation settings fall 5.00 to 13.33 points short. We will release code and trained weights.

Figures & tables

Fig. 2: Overview of CLAP. Images and language enter the pretrained vision-language backbone. The pressure reading becomes a query attending over the visual tokens, the result added back through a gated residual before the action expert predicts a block of actions. The same reading replaces the suction command in the proprioceptive state, and during execution contradicting it aborts the remaining actions and triggers a fresh inference.

Panel of Figure 2 — Fig. 3: Hardware and pressure signal. (a) The end-effector and pneumatic circuit. The module sits on the dead-end leg of the Y-fitting, so it reads the line’s static pressure rather than flow (Sec. III-A ). (b) Decoded contact state over two grasps: one seals as the cup meets the face (green), one does not (orange), leaking $4.9$ s until a retry seals. The trace is the board’s decoded regime, not a raw sensor sweep, so $\tau_{l}$ and $\tau_{h}$ are shown for reference.

Fig. 4: The supplementary segments. Drawn to scale. Top , the cup contacts the centre of the top face: (a) a pick and (b) a stacking placement, each recorded from many directions and starting positions. Bottom , the same placement with the contact point moved off centre until the rim reaches the edge, for (c)

P_{2}

and (d)

P_{4}

. Arrows give the recording direction.

Fig. 5: Reverse-order acquisition. (a) The same $P_{2}$ task recorded both ways from the head camera. The two bold labels mark the one thing the frames cannot show, which direction produced the alignment. (b) Distance from the release pose against time before release. Median and inter-quartile range, with the forward dwell shaded.

		Monochromatic						Polychromatic
Setting	Method	$P_{1}$	$P_{2}$		$P_{4}$		Avg	$P_{1}$	$P_{2}$		$P_{4}$		Avg
Setting	Method	SR	SR	Err	SR	Err	Avg	SR	SR	Err	SR	Err	Avg
Single-task	ACT [ 46 ]	30	0	N/A	0	N/A	10.00	15	0	N/A	0	N/A	5.00
	DP [ 11 ]	45	0	N/A	0	N/A	15.00	20	0	N/A	0	N/A	6.67
	DP3 [ 47 ]	20	0	N/A	0	N/A	6.67	20	0	N/A	0	N/A	6.67
	X-VLA [ 14 ]	60	65	13.29 $\pm$ 2.82	45	23.82 $\pm$ 5.38	56.67	55	60	13.94 $\pm$ 2.05	45	21.77 $\pm$ 2.87	53.33

TABLE I: Hierarchical stacking tasks. SR in percent; alignment error in mm as mean

\pm

std over successful trials, the maximum distance between corner pairs that should coincide. N/A: alignment error is not reported below a

25\%

success rate.

Variant	SR (%)			Avg	$\Delta$	Err (mm)
Variant	$P_{1}$	$P_{2}$	$P_{4}$	Avg	$\Delta$	$P_{2}$	$P_{4}$
Ours (full)	100	100	90	96.67	0.00	8.42	15.92
w/o real-time interruption	100	90	85	91.67	$-5.00$	10.79	18.48
w/o modal fusion	90	90	80	86.67	$-10.00$	10.72	17.78
w/o pressure feedback	95	85	75	85.00	$-11.67$	10.68	18.04
w/o sub-task patches	90	85	75	83.33	$-13.33$	13.39	20.35

TABLE II: Component ablation. Multi-task monochromatic setting, same protocol as Table I . Variants are defined in Sec. IV-C .

\Delta

is the drop from the full model.

Robotic Manipulation

Robotic manipulation studies how robots perceive, plan, and control physical interactions to grasp, move, assemble, and use objects.

See all 1,704 papers

VLCP: Vision Language Control Policy Closed-Loop Code Replanning for Robot Manipulation
Aug 17, 2026Dhia Naouali, Minghan Wu, Claudia Wong +2Code Generation for Robotics Robotic Manipulation
University of Monastir · St. Mildred’s-Lightbourn School · Cornell University +3
APEX: Adaptive Policy Execution for Precise Manipulation
Jun 15, 2026Mengfei Zhao, Chenxi Jiang, Tuo An +2Robotic Manipulation Adaptive Control
1MARS Lab, Nanyang Technological University
Splat2Act: From Captured Scenes to Personalized VLA Manipulation
Oct 9, 2026Xiaomeng Ye, Krishna Jaganathan, Gyubeum Lim +3Sim-to-Real Transfer Vision-Language-Action Models
Institute for Robotics and Intelligent Machines, Georgia Institute of Technology, Atlanta, GA, USA. · Waymo LLC, Mountain View, CA, USA. · Toyota Research Institute, Cambridge, MA, USA.

Aug 17, 2026J/K move · Enter open · S save

VLCP: Vision Language Control Policy Closed-Loop Code Replanning for Robot Manipulation

Dhia Naouali, Minghan Wu, Claudia Wong, Abhinav Puthran+1

University of Monastir · St. Mildred’s-Lightbourn School · Cornell University+3

Code Generation for Robotics Robotic Manipulation Robot Policy Learning Robotic Control Closed-Loop Control of VLA Models VLMs for Robotics+1

Open paper AskPDF arXiv

Turning a frontier vision-language model into a robot policy usually means fine-tuning it to emit an action representation it never saw in pretraining, which throws away much of the reasoning that made the model worth reaching for. We go the other way and keep the VLM frozen. It writes the policy as a short Python control function, with no demonstrations and no fine-tuning. Writing that code once is open-loop, though. Existing closed-loop methods react at the wrong level: they retry a fixed policy or pick a different subtask, but never rewrite the code that failed. VLCP closes the loop where the failure actually lives, on the control code, within a single episode. Every

K

steps the VLM re-observes the scene from multi-view RGB, proprioceptive state, and a state delta, then rewrites the control function from what it just saw, so a failure is caught before it compounds. We evaluate on a 57-task MuJoCo/RoboVerse sweep. This training-free policy reaches

35.1\%

pooled success, against

3.5\%

for the identical system queried once per episode. That tenfold gap holds with non-overlapping confidence intervals in every scene family. The gain traces to a

27.3\%

within-episode recovery rate on failed grasps: a miss an open-loop controller would carry to the end of the episode gets re-observed and fixed at the next replan. And the loop stays cheap. A median

84\%

of input tokens hit cache, an episode needs only about

10

compact queries, and control blocks written during any replan persist to a cross-episode skill library reused in later prompts.

Abstract

Robotic Manipulation

Explore similar work

VLCP: Vision Language Control Policy Closed-Loop Code Replanning for Robot Manipulation

APEX: Adaptive Policy Execution for Precise Manipulation

Splat2Act: From Captured Scenes to Personalized VLA Manipulation