Tensiq#
COMPANY · CURRENT WORK at Tensiq · Foundation Model Engineer
Universal tactile intelligence
for robotics.
Tensiq turns any tactile sensor into a single, physics-grounded, model-ready data format — the “MP4 for touch.” One pipeline. Any sensor. Standardized output. Designed for the next generation of dexterous robots and tactile foundation models.
What Tensiq Does#
Tactile sensors come in dozens of incompatible flavors — vision-based gels, piezoresistive arrays, multi-zone hands, custom research rigs. Every format has its own quirks, every algorithm needs its own integration, and almost none of the data ends up usable for modern foundation-model training.
Tensiq solves that. A single, sensor-agnostic pipeline turns raw tactile capture into a standardized tensor format with full physics grounding, quality auditing, and provenance — usable by real-time controllers, IL/RL pipelines, and tactile foundation models alike.
Sensor-agnostic
One pipeline, many sensors. Vision-gel, piezoresistive, multi-zone — all converge to the same standardized output.
Physics-grounded
Constitutive models compute real physical quantities — force, torque, contact state, slip — not just raw signal.
Model-ready
Output is pre-tokenized for transformers and diffusion policies. Plug into PyTorch in one line.
Auditable
Every frame carries provenance — calibration, transforms, quality flags. Full reproducibility for research & deployment.
How It Works#
A four-stage pipeline takes any vendor-specific tactile capture and emits a standardized Tensiq Tensor (TT) — ready for control loops, ML training, or long-horizon foundation models.
Capability Matrix#
| Vision-gel | Piezoresistive | Multi-zone hand | Custom rig | |
|---|---|---|---|---|
| Ingest | ✓ | ✓ | ✓ | via plugin SDK |
| Physics output | F / T / contact | F / contact / drift | per-region F / contact | configurable |
| Slip detection | ✓ | ✓ | per-region | configurable |
| Quality audit | ✓ | ✓ | ✓ | ✓ |
| ML tokenization | ✓ | ✓ | ✓ | ✓ |
Why It Matters#
One format, every sensor
No more rewriting downstream code for each new sensor vendor. The same controller, the same model, the same loader works across the catalog.
Physics, not pixels
Raw signal is great for research but ambiguous for control. Tensiq emits real forces, real torques, real contact — what robots actually need to act on.
Audit-grade by default
Every frame carries provenance. Calibration, transforms, hashes, quality flags. Pass regulatory review, reproduce a paper, deploy with confidence.
Built for foundation models
The output is pre-tokenized for transformers and diffusion policies. The tactile equivalent of what ImageNet was for vision — a substrate for the next generation of dexterous AI.
My Role#
I build the Tensiq tokenizer — the component that turns raw tactile-sensor frames into the discrete Tensiq Tensor (TT) sequences that feed multimodal foundation models. I also own the TENSIQ CLI end-to-end — pipeline execution, dataset tokenization, and cross-platform distribution — and lead the integration of tactile data into foundation models.