TACTIC: Understanding Tactile Encoders and Conditioning

01 · The question

Which design choices actually encode useful touch?

Tactile sensing is increasingly used in robotic manipulation, but the field has accumulated a wide range of encoders, training schemes, and policy-conditioning strategies. Different papers often use different tasks, environments, and evaluation protocols, making direct comparisons difficult.

TACTIC closes this gap with a shared real-world pipeline: five tactile encoders, five multimodal fusion strategies, four contact-rich tasks, and 2,180 policy rollouts.

5tactile encoders

5fusion strategies

4manipulation tasks

2,180real-world rollouts

02 · Tasks

Four tasks expose different tactile demands

Board wiping

Maintain contact and pressure while wiping a chalk line from a planar board.

Vase wiping

Control contact on a curved surface while reasoning about grasp and geometry.

USB-C insertion

Recover from small alignment errors in a precise, low-tolerance insertion task.

Screwing

Detect tightening and termination from shear and slip cues when vision is ambiguous.

03 · Models

Compare representation and conditioning choices together

Five tactile encoders

ResNet-18 baselineSARL self-supervisionUniT tactile latentsSparsh-DINO temporal featuresT³ sensor pretraining

Five fusion strategies

ConcatenationFiLM conditioningGated cross-attentionCLIP-R random alignmentCLIP-T temporal alignment

04 · Results

No single representation wins every task

Compare every policy configuration by success stage.

Board wiping success rates

Share of trials (%) · sorted by termination

05 · Robustness

Out-of-domain behavior reveals what policies attend to

Compare policies under moving disco lights and dense visual distractors.

Moving disco lights stress visual robustness during board wiping.

Disco-light success rates

Share of trials (%) · sorted by termination

06 · Conclusion

Tactile representation is task-dependent

TACTIC shows that the best encoder-fusion combination depends strongly on the task. Sparsh and T³ variants perform well on average, and simple concatenation remains competitive with more complex conditioning mechanisms.

The benchmark is designed to make these tradeoffs visible rather than hide them behind a single aggregate score.

Citation

BibTeX

@misc{bien2026tactic,
      title={TACTIC: Understanding Tactile Encoders and Conditioning for Contact-rich Robot Manipulation Policies}, 
      author={Seongjin Bien and Débora Oliveira Makowski and Carlo Kneissl and Reihaneh Mirjalili and Pankhuri Vanjani and Rudolf Lioutikov and Gitta Kutyniok and Florian Walter and Wolfram Burgard},
      year={2026},
      eprint={2609.30969},
      archivePrefix={arXiv},
}