CraftNet

Sharpa's Tactile-based Vision-Tactile-Language-Action (VTLA) Model

Author’s Note

  • CraftNet is the first commercial model to fully integrate tactile sensing into a VLA.
  • The System 0/1/2 hierarchy is similar to Figure Helix 02, but differentiated by its specialization in tactile feedback.
  • The key innovation is 100Hz high-frequency tactile control for solving the “last millimeter” problem.

Key Significance

  • First Commercial VTLA: Vision-Tactile-Language-Action, integrating tactile as a core modality
  • Three-Layer Hierarchical Architecture: System 2 (~1Hz) + System 1 (~10Hz) + System 0 (~100Hz)
  • Solves “The Last Millimeter”: High-frequency tactile feedback loop for post-contact fine manipulation
  • Synthetic Tactile Data: Enriches simulation, teleoperation, and internet video with tactile information
  • SharpaWave Integration: Combined with tactile hand featuring 1,000+ tactile pixels and 0.005N sensitivity

Sharpa CES 2026 Demo - North Humanoid with CraftNet


Overview

ItemDetails
Announced2025
CompanySharpa (Singapore)
Blogsharpa.com/blogs/news
RobotNorth Humanoid
HardwareSharpaWave Tactile Hand

CraftNet is a hierarchical Vision-Tactile-Language-Action (VTLA) model developed by Sharpa, designed for fine manipulation tasks.


Architecture: System 0/1/2

CraftNet is a hierarchical system operating across three frequency bands.

CraftNet Architecture

CraftNet Architecture: System 0/1/2 Hierarchical Structure

System 2 (Reasoning Brain) - ~1 Hz

ItemDetails
RoleTask decomposition, long-horizon planning
BaseVision-Language Model
FeatureOpen-source VLM interface
  • Decomposes human instructions into sequential sub-tasks
  • High-level reasoning and decision-making
  • Leverages open-source VLMs pre-trained on internet-scale data

System 1 (Motion Brain) - ~10 Hz

ItemDetails
RoleMotion planning, coarse action control
BaseFoundation Model
FeaturePre-contact approach optimization
  • Plans trajectories for object approach
  • Trained on public/private domain data
  • Transforms System 2 goals into executable motions

System 0 (Interaction Brain) - ~100 Hz

ItemDetails
RoleSuper high-frequency fine-motor control
BaseTactile feedback model
FeatureReal-time contact adjustment
  • Key Differentiator: Real-time tactile feedback processing
  • Continuously adjusts hand/finger positions during contact
  • Handles grasping, sliding, and complex assembly tasks

Core Technology: Tactile Integration

Limitations of Existing VLAs

Existing VLAs focus on vision-based trajectory generation with three limitations:

  1. No Tactile: Using only vision without force and tactile feedback
  2. No Post-Contact Control: Unable to handle “the last millimeter” of manipulation
  3. Unrealistic Simulation: Force/compliance patterns in simulated data don’t match reality

CraftNet’s Solutions

ProblemSolution
No TactileIntegrates force/tactile feedback alongside vision
No Post-Contact ControlSystem 0’s 100Hz high-frequency feedback loop
Data ScarcityEnriches existing data with synthetic tactile information

Data Strategy

Synthetic Tactile Data

CraftNet enriches data from various sources with tactile information:

Data SourceProcessing Method
SimulationCorrects unrealistic force/compliance patterns
TeleoperationJoint training of System 0/1 with high-quality data
Internet VideoAdds synthetic tactile information

Asynchronous Multi-Frequency Inference

  • Three systems operate independently at different frequencies
  • Temporal decoupling enables efficient computation

Hardware: SharpaWave

CraftNet is designed to work with Sharpa’s SharpaWave tactile hand.

Specifications

ItemSpec
DoF22 DoF (active)
Tactile TechnologyDynamic Tactile Array (DTA)
Tactile Pixels1,000+ per fingertip
Pressure Sensitivity0.005 N
Force Sensing6-axis
Durability1 million grip cycles
FeatureModular finger replacement

Dynamic Tactile Array (DTA)

  • “Feel by Seeing” vision-tactile fusion technology
  • Miniature camera in each fingertip
  • Handles feather-light contact to heavy load manipulation

Hardware: North Humanoid

Sharpa’s humanoid robot equipped with CraftNet.

  • Unveiled at CES 2026
  • Demonstrated fully autonomous ping-pong rallies
  • Equipped with SharpaWave hands

Company: Sharpa

ItemDetails
Founded2024
HeadquartersSingapore
R&DShanghai
BusinessMountain View, USA
AwardCES 2026 Innovation Award (Robotics)

Milestones

DateEvent
2024Sharpa founded
2025.10SharpaWave demonstrated at IROS 2025
2025.10SharpaWave mass production and shipping begins
2025.11CES 2026 Innovation Award received
2026.01North humanoid unveiled at CES 2026

Comparison with Other Hierarchical VLAs

ModelSystem 2System 1System 0Tactile
CraftNet~1Hz (VLM)~10Hz (Motion)~100Hz (Tactile)Yes
Figure Helix 02Semantic Reasoning200Hz (Visuomotor)1kHz (Balance)Yes
GR00T N110Hz (Eagle VLM)120Hz (DiT)-No

References


See Also