# DexTouch-WM: Human Tactile Pretraining

DexTouch-WM: Human Tactile Pretraining

Full article: https://haiguangboy.com/posts/dextouch-wm

## Core Method

UniTacHand has already used the MANO UV Map to unify human and robot tactile representations, and used human data together with a small number of robot demonstrations to improve manipulation. DexTouch-WM follows this path of human-to-robot tactile transfer and pushes the question further: can this experience train a world model that predicts future contact? The two connect in their goals, while their specific representations and architectures differ.

## Key Results

Human experience can transfer, but what was measured must be clearly distinguished
The human and robot pretraining tasks do not overlap; after adding human data, robot-domain contact IoU rose from 0.42 to 0.59 and F1 rose from 0.55 to 0.71.
For downstream adaptation of the world model, replacing 200 robot demonstrations with 100 human + 100 robot demonstrations raised the average correlation with real-robot policy scores from 0.646 to 0.844; this is evaluation consistency, not manipulation success rate.

## Comparison with Related Lines

- Opposing line · [An_Open_Foundation_Model_Towards](https://haiguangboy.com/posts/an_open_foundation_model_towards) `an_open_foundation_model_towards_2026_07`: The authors argue that after sharing a contact interface and action space, human touch can become a scaling axis for robot prediction contradicts joint co-training on heterogeneous human-robot data is a structurally suboptimal solution
- Opposing line · [What is closed is the kinematic loop, not the physical loop](https://haiguangboy.com/posts/video2robo) `deng_video2robo_3dgs_based_synthetic_data_from_one_video_enables_scalable_robot_2026_08`: Policy evaluation credibility and the value of synthetic training data must be validated separately contradicts the line's bet: use kinematic consistency + 3DGS photorealism to bypass physical simulation
- Same line · [Agile-WAM: Lightweight visuotactile joint prediction](https://haiguangboy.com/posts/agile-wam) `agile_wam_an_agile_tactile_world_action_model_for_contact_rich_robot_control_2026_09`: Video and tactile experts jointly predict, with tactile prediction relative to the change from the initial state validates the authors' claim: in multimodal prediction, time horizons should match the rate of change of each modality
- Same line · [N0-Foundation: Opening a new era of touch](https://haiguangboy.com/posts/mathcaln_0_foundation_towards_the_age_of_tactile_intelligence) `mathcaln_0_foundation_towards_the_age_of_tactile_intelligence_2026_09`: N0-Foundation: Opening a new era of touch
- Same line · [OpenWAM: The right answer changes with scale](https://haiguangboy.com/posts/openwam_an_open_modular_exploration_towards_systematic_world_action_model_pretra) `openwam_an_open_modular_exploration_towards_systematic_world_action_model_pretra_2026_09`: OpenWAM: The right answer changes with scale
- Same line · [PRISM: Can scale replace explicit modeling?](https://haiguangboy.com/posts/prism_precision_and_contact_rich_real_world_industrial_skill_dataset_with_multim) `prism_precision_and_contact_rich_real_world_industrial_skill_dataset_with_multim_2026_08`: PRISM: Can scale replace explicit modeling?

## Boundaries

⚠️ Another boundary: being able to predict and evaluate does not mean being able to create training data
VLAW improves policies through closed-loop generation and filtering of successful trajectories; this paper follows recorded actions to generate future observations, and after replacing half of the real training data, the average real-robot task score of π0.5 actually dropped from 0.625 to 0.494. The protocols differ, so direct ranking is not possible, but the conclusion is clear: human touch can improve world prediction, while whether synthetic data can improve control still requires separate validation.

## Author's Judgment (not paper content; cross-paper synthesis)

💡 Connections and disagreements with other lines
· Dyna-2 also reports that adding human video benefits cross-embodiment transfer; DexTouch-WM adds evidence in the tactile direction, but it relies on dedicated gloves, motion tracking, and alignment, so it cannot be equated with ordinary video scaling.
· Video2Robo extracts skills from relative object motion and bypasses human-hand retargeting; DexTouch-WM explicitly aligns hand motion and contact. The two represent different information tradeoffs, and with different tasks, they cannot yet falsify each other.