QTac: Quantitative Robotic Grasping with Tactile Sensing

Abstract

We investigate the novel task of quantitative robotic grasping and propose QTac, a tactile-driven framework designed to grasp a user-specified number of objects. Existing vision-only approaches struggle with severe occlusions in such tasks, making tactile feedback essential for reliably verifying the in-hand quantity. However, applying this concept faces two key challenges: the inherent unreliability of single-attempt grasping paradigms, and the difficulty of integrating high-dimensional tactile signals into existing manipulation models.

QTac addresses these challenges through two key innovations: (1) a human-inspired demonstration strategy that endows the robot with iterative regrasping capabilities to rectify initial count mismatches, reducing the reliance on complex sensory feedback hardware in teleoperation; (2) a decoupled tactile perception module that translates high-dimensional tactile signals into explicit quantitative feedback, directly guiding corrective regrasping. Comprehensive real-world experiments across diverse objects validate that QTac achieves reliable quantitative grasping in dense clutter and successfully extends its quantification capability to novel objects.

Method

QTac framework overview — **Overview of the QTac framework.** A dual-mode data collection strategy, a standalone tactile CountNet, and a DiTFlow-based generative policy conditioned on the count distribution and the user-specified target.

Demonstrations

Dual-Mode Data Collection

Direct-grasping trajectories give dense contact-frame supervision for CountNet, while free-adjustment trajectories provide the regrasp-rich primitives that teach the policy to correct count mismatches — without extra sensory hardware.

Perception

Tactile CountNet

A decoupled module maps bilateral tactile depth maps to a probabilistic count distribution. Keeping perception explicit and standalone lets it transfer to novel objects without retraining.

Policy

DiTFlow Regrasping

A conditional flow-matching policy fuses the count distribution, target quantity, vision, and proprioception to synthesize corrective actions, iteratively regrasping until the achieved count matches the target.

A demonstration paradigm enabling human-inspired iterative regrasping upon detecting an incorrect quantity.
A decoupled tactile perception module translating high-dimensional signals into explicit count feedback that generalizes to novel objects.
A fully integrated framework bridging data collection and policy deployment, validated across six diverse objects in the real world.

Qualitative Results

The policy continuously performs exploratory adjustments until the explicitly decoded tactile state matches the target count. Successful quantitative grasps prior to placement are highlighted; * denotes novel objects.

Quantitative Results

**Table II.** Quantitative grasping performance across 6 tasks (MAE ↓ / Success Rate % ↑). N/A: Ours (Spec.) needs per-object direct-grasping data, not collected for novel objects T5–T6.
Task	Vision-Only		End-to-End		PC [42]		π0.5 (V) [43]		π0.5 (V+T)		Ours (Gen.)		Ours (Spec.)
Task	MAE	SR	MAE	SR	MAE	SR	MAE	SR	MAE	SR	MAE	SR	MAE	SR
T1 · Envelope	0.87	33.3	0.67	43.3	1.07	30.0	1.53	16.7	2.27	6.7	0.53	50.0	0.43	56.7
T2 · Chip	1.20	23.3	1.17	23.3	1.43	26.7	2.00	10.0	1.90	13.3	0.77	36.7	0.50	50.0
T3 · Pencil	0.93	23.3	0.83	26.7	1.10	23.3	0.87	23.3	1.33	10.0	0.70	33.3	0.43	60.0
T4 · Hook	0.70	43.3	0.50	53.3	0.63	26.7	0.80	33.3	0.60	53.3	0.40	60.0	0.23	76.7
Mean (T1–T4)	0.93	30.8	0.79	36.7	1.06	26.7	1.30	20.8	1.53	20.8	0.60	45.0	0.40	60.9
T5 · Ring *	0.87	40.0	1.07	36.7	1.80	13.3	1.47	20.0	1.27	20.0	0.63	56.7	N/A	N/A
T6 · Rod *	0.50	53.3	0.73	40.0	0.70	36.7	0.73	36.7	0.67	40.0	0.33	66.7	N/A	N/A
Mean (T1–T6)	0.85	36.1	0.83	37.2	1.12	26.1	1.23	23.3	1.34	23.9	0.56	50.6	–	–

Ours (Spec.) gives the best in-domain performance (60.9% SR, 0.40 MAE on T1–T4), while the single Generalist CountNet transfers to novel objects T5–T6 with no perceptual retraining.

Demonstration Video