Part II: Frontier Robot Manipulation

Chapter 6: Hands and Touch — The Last Bottleneck in Manual-Work Automation

Written: 2026-06-08 Last updated: 2026-06-08

The final bottleneck in manual-work automation is the hand. Humans regulate grip through tactile events such as contact, slip, and release [1]. Robots still struggle with this. That is why tasks that look easy to workers remain difficult in cosmetics, food, apparel, and electronics manufacturing.

Figure 6.1: Diffusion Policy as an action-sequence generator for manipulation. source: S3 reused figure
Figure 6.1: Diffusion Policy as an action-sequence generator for manipulation. source: S3 reused figure

6.1 VLA Is Not Enough

RT-2, OpenVLA, Octo, and pi0 connect language, vision, and action, but contact-rich tasks need force and tactile feedback. A robot handling a soft tube must know not only where the tube is, but how hard it is squeezing.

6.2 The Rise of Tactile Data

GelSight, DIGIT, ReSkin, AnySkin, Sparsh, Tactile-VLA, and ForceVLA are pulling touch into robot learning [2] [3] [4]. NVIDIA's reference humanoid including tactile five-finger hands is therefore not cosmetic; it points to the real bottleneck [5].

References

  1. Roland Johansson and J. Randall Flanagan (2009). Coding and Use of Tactile Signals from the Fingertips in Object Manipulation Tasks. Nature Reviews Neuroscience.
  2. Wenzhen Yuan et al. (2017). GelSight: High-Resolution Robot Tactile Sensors. Sensors.
  3. Mike Lambeta et al. (2020). DIGIT. arXiv.
  4. Raunaq Bhirangi et al. (2021). ReSkin. arXiv.
  5. NVIDIA (2026). NVIDIA Announces Isaac GR00T Reference Humanoid Robot. NVIDIA Investor Relations.