# Hand pose is mixed into glove tactile readings

Hand pose is mixed into glove tactile readings

Full article: https://haiguangboy.com/posts/tactile-glove-pose-artifacts

## Core method

↳ ① Residual learning: force estimation is written as "pure tactile estimate + pose-induced residual," preserving the causal main path from touch to force, with pose only making corrections. The authors explicitly mention trying to directly concatenate pose and tactile features, which performed worse than pure tactile input—the model learned "pose-driven force" rather than "pose-driven artifacts."
↳ ② Supervise the auxiliary decoder only under zero load. When the hand is in the air, any change in tactile signals must come from pose, so clean labels hold automatically; once force is applied, the two become inseparable.

## Key results

· Minimum measurable force across three gloves dropped by 10.4%, 12.2%, and 18.3%, respectively; touch detection errors fell by 15%, 21%, and 24%
· Under free-moving "unseen" poses, reductions were still 10.2%, 9.6%, and 17.0%
· Fixed three failure modes: false positives in the air, detection delay at contact onset, and residual false force after release; the latter two improvements target temporal sensitivity, which is especially critical for policy training

## Related work comparison

- Same line · [T-Rex: Why tactile sensing needs separate modeling](https://haiguangboy.com/posts/t_rex_tactile_reactive_dexterous_manipulation) `t_rex_tactile_reactive_dexterous_manipulation_2026_07`: T-Rex: Why tactile sensing needs separate modeling
- Same line · [It's not "whether to have a world model," it's "whether what it outputs has structure"](https://haiguangboy.com/posts/robointer15_a_holistic_intermediate_representation_suite_for_embodied_world_mode) `robointer15_a_holistic_intermediate_representation_suite_for_embodied_world_mode_2026_07`: It's not "whether to have a world model," it's "whether what it outputs has structure"
- Same line · [Neural signals: their value and their fragility come from the same place](https://haiguangboy.com/posts/real_time_semg_based_telecontrol_of_an_assistive_robotic_arm_using_a_1d_convolut) `real_time_semg_based_telecontrol_of_an_assistive_robotic_arm_using_a_1d_convolut_2026_07`: Neural signals: their value and their fragility come from the same place
- Same line · [Tactile as a prediction target rather than an observation input](https://haiguangboy.com/posts/n_0_vtla_scaling_vision_tactile_language_action_model_with_latent_tactile_tokens) `n_0_vtla_scaling_vision_tactile_language_action_model_with_latent_tactile_tokens_2026_07`: Tactile as a prediction target rather than an observation input

## Limitations

Only the index fingertip was tested, with a single planar end-effector and normal force only; full-hand extension is future work. Each glove had only 5 participants, so the sample size was small enough to discretize p-values. More critically: the premise is that "pose is easy to obtain now," but the experiments relied entirely on high-precision optical motion capture; whether it still holds with first-person vision-level accuracy is listed by the authors as an open question.

## Author's assessment (not from the paper; cross-paper synthesis)

The judgment "don't directly concatenate auxiliary modalities into the main backbone" has now accumulated a fifth independent source in the library, spanning five levels: VLA policies, tactile-specific architectures, world model latent variables, physiological signals, and now sensor calibration. The same pitfall is repeatedly hit at completely different abstraction layers, suggesting it is not a flaw of any one model class.

This paper's value is in providing the mechanism: models tend to learn auxiliary modalities as direct predictors of the target quantity, rather than as corrections to the main path. The fix is thus clear—not concatenation, but giving it a structurally constrained position. Residual terms, prediction targets, and dedicated pathways are all different formulations of the same approach.

There is also a genuine disagreement worth noting: several tactile works in the library require specialized hardware, while this paper advocates pure algorithmic retrofitting without changing the gloves. The hardware and algorithmic routes give opposite answers to "at which layer should tactile data quality be solved," and no verdict exists yet.