IROS 2026 — MrGrasp: Multi-Rate Vision–Tactile–Language–Action for Gentle Grasping of Fragile Deformables

Venue: IROS 2026 (Pittsburgh) · paper #2149 · Huazhong Univ. of S&T (Pei, Gao, Miao, Gao, Li). Representative of: deformable-object manipulation × tactile-in-the-loop — a Vision–Tactile–Language–Action (VTLA) framework that grasps fragile deformable objects gently by running a fast tactile control loop under a slow VLM planner. Companions: Tactile VLA · Dexterous Manipulation · IROS 2026 survey.

1. Problem

Grasping fragile deformable objects is a hard contact-rich task: rapid deformation and shifting contact areas change contact state fast, and vision-only policies can't perceive/regulate contact → instability or damage. Imitation learning helps but intra-category variation makes covering all instances with demos impractical (weak generalization).

2. Method

MrGrasp is a hierarchical multi-rate VTLA framework:

• High-level VLM planner — a pretrained VLM turns semantic instruction + visual observation into domain-invariant task representations, decoupling low-level execution from appearance variability.
• Diffusion multi-rate grasp policy — a low-rate branch produces motion-trajectory priors; a high-rate branch fine-tunes grasp control in a closed loop with real-time tactile feedback, handling fast-changing contact.

3. Results

• Across diverse fragile deformables, MrGrasp beats vision-based baselines and stays robust under unseen instances and external perturbations — validating the multi-rate VTLA recipe for contact-rich fragile grasping.

4. Why it matters

MrGrasp represents two IROS 2026 gap-themes at once — deformable-object manipulation and tactile-in-the-loop control — via the now-recurring fast/slow multi-rate structure (cf. T-Rex's variable-rate MoT and [Review-Realtime-Execution]]): a slow VLM sets intent, a fast tactile loop regulates contact. It's the fragile/deformable counterpart to [T-Rex (both put touch in a high-rate reactive loop under a VLA), and evidence that tactile is becoming a first-class high-frequency channel in 2026 manipulation, not just an input.

Limitations (reviewer): grasping-scoped (not full manipulation); needs a tactile-equipped gripper; diffusion policy adds cost; evaluated on fragile-deformable grasping, not broad tasks.

5. Links

• Official program: IROS 2026 (paper #2149) · survey: IROS 2026
• Related: Tactile VLA · T-Rex · Real-Time Execution · Dexterous Manipulation

← Back to IROS 2026 survey · Home