# DexTacWAM: robot AI research model DexTacWAM (Dexterous Tactile World-Action Model) is a Sep 2026 academic visuo-tactile WAM (Yuan, Wang, Shao, Lu, Darrell, Lourentzou, Zhan; UIUC + [UC Berkeley](/companies/uc-berkeley-bair.md) + Northwestern; arXiv 2609.24976) that jointly models visual scene dynamics and distributed multi-finger contact dynamics. A finger- and pose-aware tactile compressor maps five fingertip streams into per-hand latents injected into a video diffusion [world model](/glossary/world-model.md); an action expert conditions on the joint visuo-tactile state. Author-reported mean 70.6 across six real-robot tasks versus 38.0 for the strongest baseline; compressor retains 89.4% pre-fusion contact recall with 2.26x train / 1.29x infer speedups. Research demo on a 22-DoF bimanual [Sharpa](/companies/sharpa.md) tactile platform, not a commercial product. - **Slug:** dextacwam - **Type:** Research model - **Maturity (DEPLOY ladder):** Research - **Open source:** no ## Architecture Per-finger tactile maps via frozen pretrained vision VAE + grayscale-to-RGB adapter; finger- and pose-aware 5:1 tactile compressor into per-hand latents; video diffusion DiT world model jointly denoises visual + tactile views; per-modality K/V RMS-normalized action expert (flow matching) for dexterous actions ## Key facts - **System (project-stated):** Project-stated (project-stated; also project site): DexTacWAM is a visuo-tactile world-action model (Yuan / Wang / Shao / Lu / Darrell / Lourentzou / Zhan; UIUC + [UC Berkeley](/companies/uc-berkeley-bair.md) + Northwestern; arXiv 2609.24976, Sep 21 2026) that jointly predicts visual scenes and multi-finger contact dynamics. Encodes each fingertip independently, aggregates via a finger- and pose-aware tactile compressor, and injects tactile latents into a video diffusion [world model](/glossary/world-model.md) for joint visuo-tactile world modeling. Academic research system, not a commercial product - **Six-task real-robot results (project-stated):** Project-stated (project-stated): on six contact-rich [dexterous manipulation](/glossary/dexterous-manipulation.md) tasks on a 22-DoF bimanual platform (20 real-robot trials per method per task), DexTacWAM averages 70.6 versus 38.0 for the strongest baseline (RDP). Cap: author-reported lab results; not independent third-party verification - **Tactile compressor retention / speedups (project-stated):** Project-stated (project-stated): finger- and pose-aware tactile compressor retains 89.4% of pre-fusion contact recall while enabling 2.26x faster training and 1.29x faster inference. Ablation: removing tactile world modeling drops four-task mean from 74.7 to 26.6 with the same tactile features and action expert. Cap: paper-reported metrics - **Tactile world-model ablation:** The project page says removing tactile world modeling drops the four-task mean from 74.7 to 26.6 while keeping the same tactile features and action expert (DexTacWAM). - **Sharpa 22-DoF platform:** The paper says the six-task evaluation used a bimanual platform with multi-finger 22-DoF [Sharpa](/companies/sharpa.md) tactile sensors (arXiv 2609.24976). - **UIUC, Berkeley, and Northwestern:** The project page lists authors at the [University of Illinois Urbana-Champaign](/companies/university-of-illinois-urbana-champaign.md), the [University of California, Berkeley](/companies/university-of-california-berkeley.md), and Northwestern University (DexTacWAM). ## Sources (4) 1. **DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous Manipulation (arXiv:2609.24976)** · https://arxiv.org/abs/2609.24976 · 2026-09-21 2. **DexTacWAM PDF (arXiv:2609.24976)** · https://arxiv.org/pdf/2609.24976 · 2026-09-21 3. **DexTacWAM project site** · https://dextacwam.github.io/ · 2026-09-21 4. **Secondary summary: DexTacWAM 70.6% avg across six real-robot tasks (robotsdigest)** · https://x.com/robotsdigest/status/2105059594288681337 · 2026-09-29 ## Common questions ### What is DexTacWAM? DexTacWAM (Dexterous Tactile World-Action Model) is a Sep 2026 academic visuo-tactile WAM (Yuan, Wang, Shao, Lu, Darrell, Lourentzou, Zhan; UIUC + [UC Berkeley](/companies/uc-berkeley-bair.md) + Northwestern; arXiv 2609.24976) that jointly models visual scene dynamics and distributed multi-finger contact dynamics. A finger- and pose-aware tactile compressor maps five fingertip streams into per-hand latents injected into a video diffusion [world model](/glossary/world-model.md); an action expert conditions on the joint visuo-tactile state. Author-reported mean 70.6 across six real-robot tasks versus 38.0 for the strongest baseline; compressor retains 89.4% pre-fusion contact recall with 2.26x train / 1.29x infer speedups. Research demo on a 22-DoF bimanual [Sharpa](/companies/sharpa.md) tactile platform, not a commercial product. ### Is DexTacWAM open source? No. DexTacWAM is recorded as proprietary on the DEPLOY registry. Model weights and source are not publicly available. ### What type of AI is DexTacWAM? DexTacWAM is a research model, built on a Per-finger tactile maps via frozen pretrained vision VAE + grayscale-to-RGB adapter; finger- and pose-aware 5:1 tactile compressor into per-hand latents; video diffusion DiT [world model](/glossary/world-model.md) jointly denoises visual + tactile views; per-modality K/V RMS-normalized action expert (flow matching) for dexterous actions architecture on the DEPLOY registry. ### What is DexTacWAM's maturity stage? DexTacWAM is at the research stage on the DEPLOY maturity ladder. Research stage means active development without commercial deployments on file. ### Which robots run on DexTacWAM? No robot models on the DEPLOY registry are recorded as running DexTacWAM. DEPLOY wires brain-to-model connections only when the wiring is verifiable from primary sources; absence may reflect pre-deployment or unverified manufacturer claims. _API: GET /v1/brains/449019c6-e4f8-43f0-996d-ec4d1c077b08 · canonical URL: /brains/dextacwam_