Back to feed
News Story
量子位
1 sources

First Implicit Tactile World-Action Model Being-H0.8 Released by BeingBeyond

BeingBeyond has released Being-H0.8, the first implicit tactile world-action model trained on human video data. The model integrates touch into the predict-execute-feedback loop, using over 500,000 hours of video to enable robots to anticipate contact and adjust actions in real time. In real-world experiments, it completed high-precision tasks such as retrieving objects from a bag and writing with a brush.

SynthePulse Insight · AI deep reading

Touch Enters the World Model: How Being-H0.8 'Teaches' Robots to Perceive Contact from 500,000 Hours of Video

Version 1 · 1 source

BeingBeyond releases Being-H0.8, the first model to incorporate touch into an implicit world-action model, enabling robots with a 'predict-execute-feedback' tactile closed loop through 500,000 hours of human video and an innovative data pipeline.

  • Being-H0.8 is the first implicit tactile world-action model based on human video data, learning to predict contact during pre-training and adjusting actions in real-time based on tactile feedback during execution.
  • The model uses a Mixture of Transformers (MoT) to predict interaction outcomes, with slow-fast action experts balancing global planning and real-time correction, and a universal tactile encoder unifying multi-source tactile signals.
  • The data foundation UniHand-3.0 originates from over 500,000 hours of first-person video, with TactoHand recovering contact and proximity from videos without tactile annotations, supplemented by approximately 55 hours of pressure glove data.
  • On a real dual-arm robot, it completes high-difficulty tasks that rely on touch, such as retrieving objects from a bag, writing with a brush, squeezing toothpaste, and picking up potato chips.
Open section navigationTactile Deficiency: An Open Problem for World Models

Tactile Deficiency: An Open Problem for World Models

Existing world models are mostly vision-centric, capable of observing object movement and deformation, but struggle to accurately determine whether contact occurs, where it occurs, and the magnitude of force. Meanwhile, tactile sensor data formats are inconsistent, and large-scale human video and robot datasets generally lack synchronized tactile recordings. This makes training world models with tactile capabilities difficult due to both the lack of a unified representation and the absence of scalable training data. Being-H0.8 is a solution to this problem.

Implicit Tactile World-Action Model: Architecture and Innovation

Being-H0.8 builds on Being-H0.7 by incorporating tactile modality into the latent space representation for the first time, extending it into an implicit tactile world-action model. Core modules include: Mixture of Transformers (MoT) to predict interaction outcomes; slow-fast action experts—the slow stream maintains the overall task plan, while the fast stream recalculates the next short action segment based on the latest tactile feedback; a universal tactile encoder that unifies signals from different sensors into tactile tokens; and TopoHand that aligns the action spaces of human hands, dexterous hands, and parallel grippers. During training, future tactile tokens enter the MoT posterior branch to help understand state changes; during deployment, MoT predicts ahead, and the fast stream corrects in real time.

Data Pipeline: From 500,000 Hours of Video to Unified Tactile Representation

The data foundation UniHand-3.0 originates from over 500,000 hours of first-person video, which undergoes filtering, deduplication, segmentation, language annotation, and distribution rebalancing, followed by hand motion recovery via MANO. A key innovation is TactoHand: using hand-object 3D geometric relationships, it predicts contact and proximity on 778 vertices from videos without tactile annotations, trained on 21 data sources totaling 30.1 million frames. Additionally, approximately 55 hours (5.4 million frames) of pressure glove data supplement real pressure information. The universal tactile encoder builds a coarse-to-fine tactile pyramid, mapping tactile signals of different granularities into a fixed number of tokens, each carrying a validity flag to distinguish 'no sensor' from 'zero reading'.

From Validation to High-Quality Utilization: The Data Strategy Shift of the Being-H Series

The Being-H series has gone through three stages: from validating the usability of human video (H0→H0.5), to scaling up usage (H0.5→H0.7), and now to H0.8's pursuit of high-quality utilization—extracting more accurate physical training signals from massive videos through cleaning, motion recovery, tactile annotation, and representation unification. The team believes the core barrier is no longer how much video one has, but whether the entire chain of data collection, cleaning, motion recovery, tactile annotation, world modeling, and robot control can be seamlessly connected.

Credibility boundary

This article is primarily based on QuantumBit's report on BeingBeyond's announcement, which is promotional in nature. Some performance descriptions (e.g., 'largest domestic scale') are self-proclaimed by the source and have not been independently verified. Model capabilities (e.g., completing high-difficulty tasks) are claimed by the team based on experiments, without third-party replication or comparison benchmarks.

Insight takeaway

Being-H0.8, through its innovative implicit tactile world model architecture and data pipeline of over 500,000 hours, integrates touch into the world model's predict-execute-feedback closed loop for the first time, offering a new paradigm for fine robotic manipulation. However, its actual generalization ability and data quality still require independent verification.

Sources for this version

  1. 世界模型有触觉了！50万小时视频，训出首个隐式触觉世界动作模型量子位

Primary report

量子位

Primary source