DeCAL: Towards Physically-Grounded Dexterous Vision-Language-Action Models via Contact-Aware Latent Co-Imagination
Yankai Fu, Ning Chen, Junkai Zhao, Heng Zhang, Guocai Yao, Pengwei Wang, Zhongyuan Wang, Shanghang Zhang
cs.RO · cs.AI
Quick verdict
Claims SOTA dexterous VLA via tactile gating + latent co-imagination; real-robot gains are large but from 20-trial evals, no significance tests.
Summary
DeCAL is a vision-language-action model for contact-rich bimanual dexterous manipulation that integrates fingertip tactile sensing into a Mixture-of-Transformers architecture. Three experts — an understanding expert built on Qwen3-VL, a generation expert predicting future visual and tactile latents (visuo-tactile latent co-imagination), and an action expert using Factorized Flow Matching that decouples arm and hand denoising — exchange information through a unidirectional blockwise attention mask. An adaptive visuo-tactile fusion injects deform-map tactile tokens via cross-attention scaled by a contact-aware gate. Initialized from InternVLA-A1-3B and trained on 600 teleoperated demonstrations (100 per task) on dual UR5 arms with 22-DoF SharpaWave hands and ten fingertip tactile sensors, DeCAL reports a 71% average success rate and 83.4% progress success rate across six real-world tasks, about 15 points above the strongest baseline (DECO, 55.8%), with 0.27s per action chunk on an RTX 4090. Generalization tests (four OOD settings on Twist Cap) and component ablations support each module. Main limitations: only 20 evaluation trials per task with no statistical testing or variance reporting; ablations and OOD cover one or two tasks; no code or data release is stated; the platform relies on proprietary hardware.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- DeCAL unifies understanding, imagination, and action in one Mixture-of-Transformers VLA whose three experts share information through a unidirectional blockwise attention mask.
- Adaptive visuo-tactile fusion with a contact-aware gate injects tactile information only when contact is relevant, avoiding interference with visual grounding (supported by Fig. 6 gate traces and the InternVLA-A1t comparison).
- Visuo-tactile latent co-imagination — joint prediction of future visual latents and tactile latents decoded to forces, deform maps, and raw images — supplies implicit physical world knowledge that improves contact-rich control (Table 2; Fig. 7).
- DeCAL achieves a 71% average success rate and 83.4% progress success rate across six real tasks, over 15 SR points above the strongest baseline, with strong OOD generalization (75% SR under unseen object on Twist Cap).
Strengths
Comprehensive real-robot evaluation with a purpose-built tactile control
Six bimanual contact-rich tasks with 100 demos and 20 trials each (Sec. 5.1, Table 1) against five baselines, including InternVLA-A1t — a tactile-augmented InternVLA-A1 with matched input modalities (Appendix F). The pooled margin is large: 70.8% avg SR for DeCAL vs 55.8% (DECO) and 39.2% (InternVLA-A1t).
Interpretable analyses of the tactile pathway
Beyond raw success rates: t-SNE clustering of global tactile tokens by contact mode (Fig. 5), gate values rising specifically during contact phases (Fig. 6), and predicted vs ground-truth 6-DoF fingertip force alignment across ten fingers (Fig. 7) directly support the gating and co-imagination mechanisms.
Stage-wise granularity localizes where the gain comes from
Real-time feasibility and candid limitations
Weaknesses
No statistical uncertainty anywhere
Ablations and OOD cover a fraction of the task suite
Baseline budgets heterogeneous and author-controlled
Largest ablation effect belongs to the least novel component
Low reproducibility
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
DeCAL (arXiv 2609.09119) is a method paper in robot learning proposing a unified Vision-Tactile-Language-Action model for contact-rich, bimanual dexterous manipulation on real hardware. Three claimed contributions structure the paper: (1) a Mixture-of-Transformers design with three experts — understanding (Qwen3-VL), generation (predicting future visual latents in Cosmos VAE space and tactile latents decoded to 6-DoF forces, deform maps, and raw images), and action (flow matching) — connected by a unidirectional blockwise attention mask (Fig. 3); (2) Adaptive Visuo-Tactile Fusion, where local tactile tokens enter each transformer block via cross-attention scaled by a gate derived from a global tactile token (Eq. 2); (3) Visuo-Tactile Latent Co-Imagination, i.e., joint future latent prediction (Sec. 4.4). A fourth, less novel ingredient, Factorized Flow Matching, decouples arm and hand denoising. Headline results: 71% average SR and 83.4% PSR over six real tasks (Table 1), ~15 SR points above the strongest baseline DECO (55.8%), 0.27s per 50-step action chunk on one RTX 4090, and OOD robustness on Twist Cap (75% SR under unseen object, Fig. 9). Every quantitative claim is real-robot; there is no simulation.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is DeCAL?
DeCAL is a vision-language-action (VLA) model for bimanual dexterous robot manipulation. It fuses fingertip tactile sensing into a Mixture-of-Transformers architecture with three experts (understanding, imagination/generation, action), uses a contact-aware gate to inject tactile signals adaptively, and predicts future visual-tactile latents. It reports a 71% average success rate on six real-world contact-rich tasks.
What robot hardware and sensors does DeCAL use?
A pair of 6-DoF UR5 robotic arms with two 22-DoF SharpaWave five-finger dexterous hands, three Intel RealSense D435 cameras (two wrist-mounted, one egocentric), and a 320×240 vision-based tactile sensor on each of the ten fingertips, all running at 30 Hz. Demonstrations were collected by human teleoperation using MetaGlove Pro gloves and VIVE trackers.
How much better is DeCAL than prior methods?
On the paper's six real tasks, DeCAL averages 71% success rate versus about 55.8% for the strongest baseline (DECO) — roughly 15 points. Each task was evaluated with only 20 trials and no statistical tests, so small per-task margins (e.g., 5%) are within noise; the pooled advantage is the more trustworthy finding.
Is DeCAL's code or data available?
The paper links a project webpage but does not state that code, demonstration data, or model weights are released. Reproduction also requires proprietary SharpaWave hands and Sharpa fingertip tactile sensors plus 600 teleoperated demonstrations.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Quantitative values embedded in figures (Fig. 1 banner numbers, Fig. 9 bar labels, Fig. 7 curves) arrive as garbled figure text; all numbers cited in this review were taken from Tables 1-6 and the body text, which extract cleanly.
- Figure 1's PSR banner values (69/91/78) cannot be reconciled with the abstract's 83.4% PSR under any obvious mapping; the abstract and Table 1-derived averages (71% SR, 83.4% PSR) were treated as authoritative.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.