Tactile World Models and Coding Agents: Literature Review

Can coding agents generate contact-rich environments whose synchronized vision, touch, and action data improves real robot learning beyond fixed tasks and domain randomization? The components are feasible individually. The literature does not establish that combining them automatically produces useful, physically trustworthy training data.

Environment and teacher generation are established by RoboGen, GenSim, GenSim2, and RoboTwin 2.0. UniVTAC already automates visuo-tactile data generation and correction. ViTacWorld already uses learned visuo-tactile rollouts for policy augmentation. The promising experimental question is whether contact-aware task discovery and independently checked data improve generalization at a matched total budget. This is an analytical recommendation, not an established novel result. [P50: RoboGen] [P51: GenSim] [P52: GenSim2] [P53: RoboTwin 2.0] [P15: UniVTAC] [P27: ViTacWorld]

The first implementation should use a calibrated pressure-array branch if FlexiTac hardware is the target, or an optical tactile branch if access to existing optical datasets and simulators is the priority. Establish data value with a small policy before committing to a large tactile VLA or generative world model.

Coverage and evidence

This targeted review contains 56 distinct papers/preprints: 12 linked works and one named method from the seed reading list, plus 43 works outside that list. The seed document has five tabs; repeated reviews and the GalaxeaVLA companion repository are not counted as additional papers. Hardware and budget links are outside the paper count.

Methods or results sections were inspected for 23 works; the remaining 33 have abstract/project-level coverage, sometimes supplemented by repository inspection. These labels describe evidence depth, not full replication or exhaustive full-paper reading. All 56 primary landing pages were retrieved. Findings reflect available sources through September 8, 2026; dates in the catalog are first-submission dates where available. Preprints are not assumed to be peer reviewed.

Area	Distinct works	Decision supported
Policy and fusion	14	What student and fusion baselines to train
Simulation and sensors	13	How to generate physically interpretable touch
Data and representations	12	What real data and encoders can be reused
Agents and generation	10	What generation mechanisms already exist
World models	7	What to predict and how to evaluate it

Seed reading list: item-by-item coverage

Seed work	Role in this project	Coverage
3D-ViTac	Spatially aligns tactile measurements and visual geometry for fine manipulation.	Methods/results
VT-Refine	Uses real demonstrations and calibrated tactile simulation to fine-tune bimanual assembly policies.	Methods/results + code
Policy Consensus	Combines modality-specific policies through agreement to improve manipulation.	Methods/results
PhysTwin	Reconstructs deformable objects with a physics-based representation and visual appearance.	Methods/results
FELT	Learns tactile signal generation from paired visual and tactile data for manipulation.	Methods/results
Beyond Teleoperation	Uses explicit kinematic retargeting of human demonstrations to improve VLA robustness.	Methods/results
OpenVLA	Open VLA training and adaptation reference.	Abstract/project
G0.5	Combines robot reasoning and action in an autoregressive stream.	Abstract/project
ASPIRE	Discovers, repairs, and reuses executable robot skills with an agent.	Methods/results + code
ENPIRE	Improves robot code policies through real execution, verification, and resets.	Methods/results + code
SimFoundry	Builds modular simulation scenes for policy learning and evaluation.	Methods/results + code
MolmoAct2	Action reasoning model with real-robot deployment workflows.	Abstract/project + code
ACT / ALOHA	Introduces action chunking for fine-grained bimanual imitation learning with low-cost hardware.	Abstract/project

ACT was named without a paper link and is resolved to the ALOHA/ACT paper. G0.5 and its GalaxeaVLA repository refer to one work. The literature-review tab repeats 3D-ViTac, VT-Refine, FELT, Policy Consensus, and PhysTwin; they remain single entries. FlexiTac hardware is named in the planning material; its separate sensor paper is included in the expanded set.

Yunzhu Li and FlexiTac research lineage

Three related lines should guide the reading order. First, 3D-ViTac and FlexiTac address accessible tactile hardware and representations that connect local contact to visual geometry. VT-Refine adds a real-to-simulation-to-real training path. Together, they provide the most relevant starting point for pressure-array manipulation. [P01: 3D-ViTac] [P14: FlexiTac] [P02: VT-Refine]

Second, RoboPack shows how tactile history can inform structured dynamics and hidden physical properties for planning. It offers a useful alternative to spending the first experiment on photorealistic tactile video generation: a compact contact-informed state may already improve decisions. [P25: RoboPack]

Third, PhysTwin and Interactive World Simulator address different forms of world construction: explicit deformable reconstruction and learned action-conditioned visual simulation. Their relevance is complementary. Neither should be treated as evidence that a visual world simulator already supplies calibrated FlexiTac observations. [P04: PhysTwin] [P31: Interactive World Simulator]

Recommended sequence: 3D-ViTac → VT-Refine → RoboPack → Interactive World Simulator, with FlexiTac as the hardware reference. Read the closest external work, UniVTAC and ViTacWorld, alongside this sequence so that the project is positioned against both the lab lineage and competing approaches.

Sensor simulation and platform choice

Pressure arrays, optical tactile images, marker displacement fields, and wrist force/torque are different observations. A dense-looking image is not evidence of a dense force measurement. Optical sensor simulation includes elastomer deformation and the optical response; a pressure-array simulator needs an appropriate mapping from contact to taxel readings. Encoders and datasets do not transfer automatically between these sensing principles. [P14: FlexiTac] [P17: Taxim] [P20: TacEx] [P34: ForceVLA2]

Path	Relevant work	Practical implication
FlexiTac / pressure arrays	[P01: 3D-ViTac] [P02: VT-Refine] [P14: FlexiTac]	Start from a sensor calibration and a rigid insertion or grasp-adjustment task. Keep array geometry and units explicit.
Optical rendering	[P16: TACTO] [P17: Taxim] [P18: Tacchi]	Useful sensor models and rendering baselines; robot dynamics, teachers, and task success still require integration.
Optical simulation at scale	[P19: TacSL] [P20: TacEx] [P21: Taccel] [P15: UniVTAC]	TacSL, TacEx, Taccel, and UniVTAC offer different engine and contact-model choices; benchmark the actual sensor/task configuration.
Differentiable or difficult contacts	[P22: DiffTactile] [P24: IsaacIPC]	DiffTactile supports calibration and control optimization; IsaacIPC explicitly targets contact-pressure resolution as well as rendering.
Fixed task evaluation	[P23: ManiFeel]	ManiFeel compares tactile representations and policy architectures, making it a useful benchmark for task-selection claims.

Taccel uses IPC with affine-body dynamics to simulate soft tactile pads and robotic systems on GPUs. Its published throughput is tied to specified scenes and hardware; it is not a universal speed estimate. IsaacIPC adds a contact-surface potential aimed at improving pressure distributions, which is particularly relevant when the output label is a contact field. Both deserve evaluation before writing a new contact solver. [P21: Taccel] [P24: IsaacIPC]

ManiFeel uses Isaac Gym and TacSL, with tactile images and normal/shear force-field representations. Its real-world policies are trained on real demonstrations to test whether modality trends agree with simulation. Those results support comparative benchmarking, not direct transfer of a simulation-trained policy. [P23: ManiFeel]

Recommended calibration gate: compare simulated and measured indentation, contact onset, loading/unloading, grasp force distribution, and time alignment before collecting training trajectories. Where a sensor cannot directly observe shear or texture, avoid assigning it the corresponding optical-sensor supervision. This is a proposed validation procedure.

World models: seven distinct roles

Work	Model/data role	Boundary that changes the experiment
[P25: RoboPack]	Tactile-informed structured dynamics for packing and planning.	Test latent physical-state estimation before assuming pixel prediction is necessary.
[P26: VT-WM]	Multi-task visual/tactile latent prediction and planning.	DIGIT 360; planning evaluation uses the same scene and objects. Unseen physical-property transfer remains open.
[P27: ViTacWorld]	Action-conditioned paired rollouts for training augmentation and policy evaluation.	Very close prior art; account for real pretraining, downstream rollout adaptation, and selection of generated samples.
[P28: Dream-Tac]	Joint future visual/tactile and action modeling.	A world-action baseline; adding tactile prediction alone is insufficient differentiation.
[P29: ContactWorld]	Study of representations for contact-rich world models.	Control representation choices when testing the effect of generated data.
[P30: OmniVTA / OmniViTac]	OmniViTac data plus predictive contact modeling and corrective control.	Large real-data baseline; a learned contact model can support fast correction without rendering all future visual frames.
[P31: Interactive World Simulator]	Interactive action-conditioned visual simulation for policy training/evaluation.	Visual baseline in the Yunzhu Li lineage; tactile prediction is not established in the inspected evidence.

VT-WM is especially useful for designing evaluation: it separates object persistence and contact-consistent motion from downstream planning. Its appendix explicitly limits object and scene generalization, and notes the cost of sampling rollouts for planning. This leaves room for a carefully defined evaluation on new geometries and physical parameters, but not for claiming that tactile world modeling itself is new. [P26: VT-WM]

For the first world model, predict contact onset, region, calibrated magnitude, and action-conditioned changes in a compact latent state. Include failures and probing sequences; success-only demonstrations omit much of the information needed to learn dynamics. Compare against an equally sized visual model and against a direct policy with the same tactile encoder. These are proposed experiments.

Coding agents and synthetic data generation

The agent can operate at three levels: constructing tasks and scenes, producing successful controllers or demonstrations, and choosing additional data based on a student's weaknesses. These are separate interventions. A benchmark should identify which intervention causes the improvement instead of bundling all three into one comparison.

Work	Established mechanism	What remains to test here
[P50: RoboGen] [P51: GenSim]	Generated environments, task curricula, and robot-learning supervision.	Contact-specific diversity and calibrated tactile observations.
[P52: GenSim2] [P53: RoboTwin 2.0]	More complex tasks, execution-code generation, simulation feedback, and data for policy training.	Whether tactile-aware selection beats these strong visual-generation baselines.
[P54: MimicGen]	Expands seed demonstrations by adapting them to new configurations.	Whether adaptation remains valid during fine contact and hidden misalignment.
[P09: ASPIRE] [P10: ENPIRE]	Robot program discovery, execution, repair, and reuse.	Whether the resulting teachers generate transferable neural-policy data.
[P11: SimFoundry]	Modular scene construction and digital scene variations.	Physics calibration and the availability of the complete data/training pipeline.
[P55: ChronoAgentic (formerly Coding Agent Is Good As World Simulator)]	Executable physics worlds with review and deterministic diagnostics.	Calibrated tactile labels and downstream robot-learning value.
[P56: Code World Model]	Persistent executable state connected to generative visual observations.	Preserving contact geometry and force consistency through visual synthesis.

ChronoAgentic changes the novelty assessment. The paper originally titled “Coding Agent Is Good As World Simulator” has an August 2026 revision with a new title and a larger evaluation. The latest version generates executable PyChrono worlds, checks trajectories, reviews images/video, and repairs code. Thus an agent plus a physics-checking loop is already prior art. Its authors also identify judge/model overlap, differing clip lengths, and the inability of video assessment to directly certify physical quantities. The relevant gap is independently measured sensor consistency and robot learning, not just simulation plausibility. [P55: ChronoAgentic (formerly Coding Agent Is Good As World Simulator)]

Code World Model offers a different decomposition: maintain state through code, produce a proxy video, and render rich observations with a video model. This could later improve visual appearance, but it is not evidence that generated texture or motion preserves tactile contact labels. Test visual refinement with frozen geometry and sensor labels before adopting it in a synchronized training pipeline. [P56: Code World Model]

Real data, synthetic data, and representation pretraining

Real tactile data serves at least three purposes: calibrating the sensor, initializing representations, and grounding the deployment distribution. Keeping these purposes separate avoids calling a system “synthetic-only” when its tactile encoder or world model depends on substantial real interaction data.

Resource	What the source provides	How to use it
[P30: OmniVTA / OmniViTac]	OmniViTac: 21,000+ trajectories across 86 tasks in the current paper.	Action-conditioned contact pretraining and comparison to real-data scaling; confirm access and sensor mapping.
[P40: FreeTacMan]	Handheld visuo-tactile demonstrations; the project reports over 10,000 trajectories across 50 tasks.	A real-data baseline and practical collection design, not a synthetic corpus.
[P41: exUMI]	Action-aware tactile prediction and robot-free demonstrations.	Test whether temporal pretraining explains gains attributed to the world model.
[P42: TacUMI (Cheng et al.)] [P43: ViTaMIn]	Multimodal handheld collection, segmentation, and compliant-gripper demonstrations.	Inform synchronization and calibration; preserve embodiment-specific contact mechanics.
[P44: Sparsh] [P45: Sparsh-X] [P47: AnyTouch 2]	Self-supervised tactile encoders and dynamic/multisensory tactile benchmarks.	Compare pretrained and from-scratch encoders at the same downstream data budget.
[P46: UniTouch] [P48: Octopi] [P49: Touch and Go]	Multimodal alignment, object-property reasoning, and paired human vision/touch observations.	Useful perception pretraining; these labels do not automatically define robot actions.

OmniVTA's tactile representation uses marker displacement rather than treating every sensor as an RGB image; its control architecture compares predicted and measured contact signals. This provides a concrete contact-representation and control precedent. OmniViTac is the dataset accompanying OmniVTA, so the pair counts as one paper here. [P30: OmniVTA / OmniViTac]

Names require care. “TacUMI” refers to Cheng et al.'s standalone paper, and also appears as the name of collection devices in TouchGuide and OmniVTA. These systems should not be merged on name alone. Record the paper, hardware, calibration, action definition, and actual dataset identifier. [P42: TacUMI (Cheng et al.)] [P36: TouchGuide] [P30: OmniVTA / OmniViTac]

For generated sequences, store aligned visual observations, raw and calibrated tactile values, sensor validity, proprioception, executed actions, timestamps, controller rate, instruction, contact events, and termination reason. Record task/asset lineage and whether touch is measured, physics-simulated, or predicted from vision. Preserve failed contact and recovery segments for world-model training. These are proposed dataset requirements.

Policy and VLA baselines

A fair experiment needs competent tactile integration. Comparing an elaborate method only to vision-only control or naive concatenation cannot isolate the quality of synthetic data. Policy Consensus, TacFiLM, VLA-Touch, and TouchGuide offer different fusion and adaptation strategies that should influence baseline selection. [P03: Policy Consensus] [P33: TacFiLM] [P39: VLA-Touch] [P36: TouchGuide]

TacFiLM integrates pretrained tactile features into visual features and adapts OpenVLA-OFT with LoRA. The experiments use DIGIT and begin insertion with the object already grasped. This is a useful strong fusion baseline, with a narrower setting than full grasp-and-insert manipulation. A pressure-array version requires a suitable tactile encoder. [P33: TacFiLM]

Reactive Diffusion Policy and OmniVTA distinguish slow planning from fast contact response. Tactile-VLA also includes force-aware control and demonstrations of failures and corrective reasoning. If the task requires a rapid reaction, increasing the number of tactile tokens in a slowly executed action chunk may not address the bottleneck. [P37: Reactive Diffusion Policy] [P30: OmniVTA / OmniViTac] [P32: Tactile-VLA]

FACTR already uses a curriculum to discourage over-reliance on vision and increase use of force. Therefore a modality-aware training curriculum is not new in the broad sense. The proposed contribution would need to establish something more specific, such as generating novel contact interventions and improving transfer beyond a fixed-task curriculum. [P38: FACTR]

Start with ACT or a diffusion policy and one stronger fusion baseline. Add a modular touch-guidance model if inexpensive correction is the goal. Evaluate OpenVLA, MolmoAct2, or G0.5 only after the data pipeline improves a small policy; these backbone references do not themselves validate native pressure sensing. [P13: ACT / ALOHA] [P07: OpenVLA] [P12: MolmoAct2] [P08: G0.5]

Proposed generation and verification pipeline

Task specification → executable environment → teacher interaction → synchronized observations → independent acceptance → training. The agent proposes tasks and controllers; a fixed simulator and calibrated sensor model generate observations. A separate acceptance process decides which trajectories can enter training.

The task specification should distinguish visible information from hidden contact state. For example, vision identifies a target socket while the held object's small misalignment must be inferred through touch during contact. The teacher approaches visually, probes or adjusts after contact, and retreats when necessary. Crucially, the selected pressure array must actually make the hidden state observable; an impossible sensing problem is not a useful multimodal benchmark.

Acceptance needs four kinds of evidence. First, the environment executes and remains numerically stable. Second, geometric contact and sensor readings agree under calibrated conditions. Third, the teacher succeeds according to a success function outside the agent's editable code. Fourth, the data teaches a deployable student something useful on independent development cases. A compelling video or a high teacher success rate is insufficient by itself.

Privileged state may be used by teachers, calibration, and evaluators. It must be excluded from student observations unless the real robot can measure it. Record these channels explicitly to make leakage detectable. Freeze final test assets, physical conditions, and scoring rules before the agent starts choosing training tasks.

Candidate contributions and existing overlap

Candidate claim	Closest overlap	More precise experiment
Coding agents generate robot environments	[P50: RoboGen] [P51: GenSim] [P52: GenSim2] [P53: RoboTwin 2.0]	Compare contact-aware generation with the same agent and compute budget using generic task generation.
Automatically generate visual and tactile data	[P15: UniVTAC] [P27: ViTacWorld]	Establish calibrated pressure-array data generation or stronger transfer at controlled real-data cost.
A physics critic repairs generated worlds	[P55: ChronoAgentic (formerly Coding Agent Is Good As World Simulator)] [P15: UniVTAC]	Use independently measured contact/sensor consistency and show it predicts downstream data value.
Tactile predictions improve control	[P25: RoboPack] [P26: VT-WM] [P28: Dream-Tac] [P30: OmniVTA / OmniViTac]	Separate the benefit of predictive contact state from encoder pretraining and fast feedback.
Training teaches the model to use touch	[P38: FACTR] [P03: Policy Consensus] [P33: TacFiLM]	Generate appearance-matched contact interventions and compare against fixed-task modality curricula.
Synthetic data improves real robots	[P02: VT-Refine] [P27: ViTacWorld] [P31: Interactive World Simulator] [P53: RoboTwin 2.0]	Report a scaling curve and real transfer under held-out geometry/physics with total generation cost.

Most defensible initial hypothesis: under a fixed real-data and generation budget, agent-selected contact interventions produce more useful visuo-tactile training data than fixed templates or generic agent-generated tasks. The evidence should show both physical consistency and improved transfer. This is a candidate hypothesis, not a claim of firstness.

Experiment plan and decision gates

Experiment	Controlled comparison	Outcome required
Sensor validity	Known indentation and grasp sequences; timestep and parameter changes.	Stable contact timing, regions, and calibrated magnitude on held-out measurements.
Modality usefulness	Vision + proprioception; touch + proprioception; fused observations; available force/gripper feedback.	Fused policy gains on independent cases, with sensing adequacy and capacity controlled.
Data generation	Fixed templates + randomization; generic coding agent; contact-aware coding agent.	Compare equal valid-data volume and equal total generation budget; include rejected attempts in cost.
Verification	Identical proposed tasks with different acceptance criteria.	Determine whether stronger contact checks improve held-out policy success rather than merely reject more data.
Fusion and timing	Concatenation vs one strong fusion method; slow policy vs fast correction where relevant.	Show data gains persist beyond a weak fusion or controller baseline.
World-model value	Vision-only dynamics; visual/tactile dynamics; direct tactile policy.	Measure contact prediction and planning/recovery, not only image similarity.
Real transfer	Fixed calibration and real demonstration budgets; novel objects and contact conditions.	Report success, recovery, load violations, and uncertainty across independent runs.

Use paired interventions to distinguish the modalities: keep appearance similar while varying contact state, and keep contact similar while changing the visually specified target. Zeroing or shuffling touch is a diagnostic but can introduce out-of-distribution artifacts, so it should not be the sole evidence that tactile sensing is necessary.

Group train/test splits by asset family and task lineage, not individual frames. Report several independent seeds and uncertainty over task/asset groups. Count agent calls, simulation time, rejected environments, teacher failures, and real calibration effort. If an agent repeatedly selects cases that fool one student, validate with a second policy family or independent development split.

The first deliverable should be one reproducible rigid contact task, a synchronized dataset containing success and recovery, and a transfer comparison against fixed randomization. Expand to deformables only after pressure consistency and data value are established. Without real hardware, label the conclusions simulation-only.

Priority reading order

Order	Read together	Question to resolve
1	[P02: VT-Refine] [P15: UniVTAC]	How much of the calibrated tactile generation pipeline already exists?
2	[P27: ViTacWorld] [P30: OmniVTA / OmniViTac]	What real data and selection are required for world-model-generated training data?
3	[P52: GenSim2] [P53: RoboTwin 2.0] [P55: ChronoAgentic (formerly Coding Agent Is Good As World Simulator)]	What exactly can the agent contribute beyond existing code-generation and repair systems?
4	[P23: ManiFeel] [P38: FACTR] [P33: TacFiLM]	How do task choice, modality curriculum, and fusion affect the baseline?
5	[P25: RoboPack] [P26: VT-WM]	Would a structured or latent contact model solve the first research question?
6	[P40: FreeTacMan] [P41: exUMI] [P14: FlexiTac]	What real calibration and demonstration collection is practical for the selected sensor?

Searchable paper catalog

Paper	Area / provenance	Evidence	Sensing / data	Contribution	Boundary / use
P01 · 3D-ViTac 2024/10/31 Full citation	Policy and fusion Linked in seed Core	Methods/results	Pressure array + 3D vision	Spatially aligns tactile measurements and visual geometry for fine manipulation.	Baseline for FlexiTac representations; array design and calibration must match the actual hardware.
P02 · VT-Refine 2025/10/16 Full citation	Simulation and sensors Linked in seed Core	Methods/results + code	Pressure array + vision	Uses real demonstrations and calibrated tactile simulation to fine-tune bimanual assembly policies.	Closest pressure-array transfer precedent; rigid CAD tasks do not establish general deformable or synthetic-only VLA learning.
P03 · Policy Consensus 2025/09/27 Full citation	Policy and fusion Linked in seed Core	Methods/results	Pressure array in real tasks; other modalities in RLBench	Combines modality-specific policies through agreement to improve manipulation.	Use as a strong fusion control; its RLBench results are not tactile experiments.
P04 · PhysTwin 2025/03/23 Full citation	Simulation and sensors Linked in seed Supporting	Methods/results	Video to deformable digital twin	Reconstructs deformable objects with a physics-based representation and visual appearance.	Useful later for deformable assets; visual reconstruction alone does not calibrate tactile forces.
P05 · FELT 2026/07/22 Full citation	Data and representations Linked in seed Core	Methods/results	Vision-predicted pressure	Learns tactile signal generation from paired visual and tactile data for manipulation.	Include as a pseudo-touch baseline; inferred touch cannot supply independent measurements of visually hidden contact.
P06 · Beyond Teleoperation 2026/08/07 Full citation	Data and representations Linked in seed Supporting	Methods/results	Human motion / vision	Uses explicit kinematic retargeting of human demonstrations to improve VLA robustness.	Relevant to demonstration conversion; tactile supervision is not established by this source.
P07 · OpenVLA 2024/06/13 Full citation	Policy and fusion Linked in seed Supporting	Abstract/project	Vision + language + actions	Open VLA training and adaptation reference.	A possible backbone, not a native pressure-array data generator.
P08 · G0.5 2026/08/12 Full citation	Policy and fusion Linked in seed Supporting	Abstract/project	Vision / reasoning / actions	Combines robot reasoning and action in an autoregressive stream.	GalaxeaVLA is its companion repository, not an additional paper; tactile support must be separately established.
P09 · ASPIRE 2026/06/30 Full citation	Agents and generation Linked in seed Core	Methods/results + code	Vision + executable robot programs	Discovers, repairs, and reuses executable robot skills with an agent.	Useful teacher-generation precedent; code-policy success is distinct from training a neural tactile policy.
P10 · ENPIRE 2026/06/18 Full citation	Agents and generation Linked in seed Core	Methods/results + code	Vision and robot feedback	Improves robot code policies through real execution, verification, and resets.	Gripper torque feedback is not a spatial tactile array; repeated-attempt success must not be reported as single-attempt success.
P11 · SimFoundry 2026/06/26 Full citation	Agents and generation Linked in seed Core	Methods/results + code	Visual scene generation	Builds modular simulation scenes for policy learning and evaluation.	Scene assets help diversity; estimated physical properties require calibration and release scope is narrower than the full paper pipeline.
P12 · MolmoAct2 2026/05/04 Full citation	Policy and fusion Linked in seed Supporting	Abstract/project + code	Vision + reasoning + actions	Action reasoning model with real-robot deployment workflows.	Candidate student backbone; inspected release did not establish native spatial pressure input.
P13 · ACT / ALOHA 2023/04/23 Full citation	Policy and fusion Named in seed Supporting	Abstract/project	Vision + proprioception + actions	Introduces action chunking for fine-grained bimanual imitation learning with low-cost hardware.	A compact student baseline can reveal whether synthetic data helps before large VLA training.
P14 · FlexiTac 2026/04/30 Full citation	Simulation and sensors Expanded search; sensor named in seed Core	Abstract/project + code	Piezoresistive pressure arrays	Provides a low-cost scalable tactile sensing platform.	Keep separate from optical GelSight/DIGIT simulators; the IsaacSim example is not the calibrated VT-Refine pipeline.
P15 · UniVTAC 2026/02/10 Full citation	Simulation and sensors Expanded search Core	Methods/results	Optical tactile + vision + actions	Unifies visuo-tactile simulation, automated data generation, correction, learning, and benchmarking.	Direct prior art for synthetic tactile data and contact checks; adaptation to FlexiTac requires a different sensor model.
P16 · TACTO 2020/12/15 Full citation	Simulation and sensors Expanded search Supporting	Abstract/project	Optical tactile rendering	Provides a fast renderer and interfaces for simulating vision-based tactile sensors.	Useful rendering baseline; visual similarity does not by itself establish accurate contact dynamics.
P17 · Taxim 2021/09/09 Full citation	Simulation and sensors Expanded search Supporting	Abstract/project	GelSight optical response and markers	Calibrates optical rendering and models marker displacement using an example-based simulator.	Sensor rendering component rather than a complete environment and teacher generator.
P18 · Tacchi 2023/01/19 Full citation	Simulation and sensors Expanded search Supporting	Abstract/project	Optical elastomer deformation	Uses particles and Taichi for a pluggable tactile deformation simulator.	Compare deformation cost and fidelity; full robot integration and sensor calibration remain necessary.
P19 · TacSL 2024/08/12 Full citation	Simulation and sensors Expanded search Core	Abstract/project	Visuotactile images and contact-force distributions	GPU-parallel tactile simulation and policy learning with sim-to-real transfer.	Central to the VT-Refine implementation lineage; Isaac Gym and Isaac Sim interfaces are not interchangeable.
P20 · TacEx 2024/11/07 Full citation	Simulation and sensors Expanded search Core	Abstract/project	Optical tactile, soft contacts	Combines deformable contact simulation with tactile rendering in Isaac Sim.	A platform candidate for optical sensors; validate throughput and supported sensor configuration in the intended workload.
P21 · Taccel 2025/04/17 Full citation	Simulation and sensors Expanded search Core	Methods/results	Vision-based tactile with soft gel dynamics	Uses GPU IPC and affine-body dynamics for parallel robot and tactile simulation.	Strong optical physics baseline; published scaling conditions should not be assumed for a different GPU or task.
P22 · DiffTactile 2024/03/13 Full citation	Simulation and sensors Expanded search Supporting	Abstract/project	Differentiable contact and tactile signals	Differentiable physics for contact-rich manipulation with deformable sensors and objects.	Relevant to calibration and control optimization; scalability for large agent-generated corpora is unestablished here.
P23 · ManiFeel 2025/05/24 Full citation	Simulation and sensors Expanded search Core	Methods/results	Multiple visuotactile representations	Benchmarks supervised visuotactile policies across contact-rich and visually challenging tasks.	Real-world policies were trained on real datasets to compare modality trends; these experiments are not direct simulation-to-real policy transfer.
P24 · IsaacIPC 2026/05/23 Full citation	Simulation and sensors Expanded search Supporting	Abstract/project	Contact physics + rendering	Couples GPU IPC with Isaac Sim/Lab and introduces a tactile-surface contact potential to improve contact-pressure distributions.	Candidate for difficult contacts and deformation, not proof that arbitrary generated tasks transfer to real hardware.
P25 · RoboPack 2024/07/01 Full citation	World models Expanded search Core	Methods/results	Soft-Bubble tactile + visual geometry	Learns tactile-informed dynamics and latent physical information for dense packing and planning.	Yunzhu Li research lineage; structured latent physics is an alternative to generating tactile video frames.
P26 · VT-WM 2026/02/05 Full citation	World models Expanded search Core	Methods/results	DIGIT 360 tactile images + vision	Learns multi-task visuo-tactile predictive models and uses imagined rollouts for planning.	Planning uses the same scene and objects; generalization to new physical properties remains open. CEM planning is computationally costly.
P27 · ViTacWorld 2026/07/24 Full citation	World models Expanded search Core	Methods/results	Optical tactile + vision + actions	Uses real and simulated interactions to learn action-conditioned visuo-tactile rollouts for policy augmentation and evaluation.	Very close to the proposed data-generation goal; real-rollout adaptation and generated-data selection remain material assumptions.
P28 · Dream-Tac 2026/06/07 Full citation	World models Expanded search Core	Methods/results	Optical tactile + vision + actions	Unifies future visual/tactile prediction with action modeling.	Predicting touch inside a world-action model is established prior art; compare downstream transfer and failure recovery.
P29 · ContactWorld 2026/06/11 Full citation	World models Expanded search Core	Methods/results	Vision / point clouds and tactile fields	Studies representation choices for contact-rich visuo-tactile world modeling.	Use to separate benefits of data generation from benefits of representation design.
P30 · OmniVTA / OmniViTac 2026/03/19 Full citation	World models Expanded search Core	Methods/results	Optical tactile / displacement fields + vision + actions	Combines a large visuo-tactile-action dataset, predictive contact modeling, fusion, and fast corrective control.	OmniViTac is the dataset within this paper, not another paper; real data scale and embodiment need separate accounting.
P31 · Interactive World Simulator 2026/03/09 Full citation	World models Expanded search Core	Abstract/project	Action-conditioned visual prediction	Yunzhu Li and collaborators study learned simulation for policy training and evaluation.	Relevant visual world-model baseline; tactile sensing is not established by the reviewed project material.
P32 · Tactile-VLA 2025/07/12 Full citation	Policy and fusion Expanded search Core	Methods/results	Normal/shear tactile feedback + vision + language	Connects VLA physical knowledge and tactile reasoning with hybrid position-force control.	Demonstrations include tactile-guided forces and failures; FlexiTac normal pressure does not directly reproduce the reported normal/shear sensing.
P33 · TacFiLM 2026/03/15 Full citation	Policy and fusion Expanded search Core	Methods/results	Pretrained optical tactile features + vision	Modulates intermediate visual features with tactile representations for lightweight VLA adaptation.	Experiments use DIGIT and OpenVLA-OFT with LoRA; insertion starts with the object already held. Pressure-array compatibility remains untested.
P34 · ForceVLA2 2026/03/16 Full citation	Policy and fusion Expanded search Supporting	Abstract/project	Force awareness + hybrid position-force actions	Builds force-aware VLA control for contact-rich tasks.	Include a force-feedback control baseline; force/torque sensing is distinct from dense tactile arrays.
P35 · OmniVTLA 2025/08/12 Full citation	Policy and fusion Expanded search Supporting	Abstract/project	Optical and force-based tactile	Uses semantic-aligned tactile representations and introduces the ObjTac dataset.	Cross-sensor fusion reference; object/property samples are not automatically full action-labeled trajectories.
P36 · TouchGuide 2026/01/28 Full citation	Policy and fusion Expanded search Core	Abstract/project	Tactile-guided action refinement	Uses a contact model to guide diffusion/flow policy actions at inference time.	Tests whether synthetic contact data helps a small corrective model without retraining an entire VLA.
P37 · Reactive Diffusion Policy 2025/03/04 Full citation	Policy and fusion Expanded search Core	Abstract/project	Visual planning + fast tactile/force feedback	Combines slower action-chunk planning with fast contact-responsive control.	A temporal control baseline; action chunking alone can miss contact changes during execution.
P38 · FACTR 2025/02/24 Full citation	Policy and fusion Expanded search Core	Abstract/project	Robot force feedback + vision	Uses a visual corruption curriculum to encourage policy attention to force.	Critical comparison for a proposed modality curriculum; this work does not generate new environments.
P39 · VLA-Touch 2025/07/23 Full citation	Policy and fusion Expanded search Core	Abstract/project	Tactile semantics + low-level touch control	Adds high-level tactile feedback and diffusion-based action refinement around a frozen VLA.	A modular integration baseline; distinct from Tactile-VLA despite similar names.
P40 · FreeTacMan 2025/06/02 Full citation	Data and representations Expanded search Core	Abstract/project	Robot-free optical visuotactile demonstrations	Collects contact-rich demonstrations with a handheld gripper and supports tactile pretraining.	Real-data calibration and co-training source; robot-free collection does not mean synthetic data.
P41 · exUMI 2025/09/18 Full citation	Data and representations Expanded search Core	Abstract/project	9DTact + pose / gripper motion	Extends UMI and pretrains tactile representations through action-aware temporal prediction.	Direct alternative for obtaining action-conditioned tactile data; check action and coordinate conversion across embodiments.
P42 · TacUMI (Cheng et al.) 2026/01/21 Full citation	Data and representations Expanded search Supporting	Abstract/project	ViTac + force/torque + pose	Synchronizes multimodal handheld demonstrations and segments contact-rich sequences.	Name collision: this paper differs from the TacUMI devices described by TouchGuide and OmniVTA.
P43 · ViTaMIn 2025/04/08 Full citation	Data and representations Expanded search Supporting	Abstract/project	Compliant Fin Ray gripper + visuotactile sensing	Robot-free demonstration collection and multimodal representation learning.	Its compliant gripper changes contact mechanics; demonstration transfer needs more than matching camera views.
P44 · Sparsh 2024/10/31 Full citation	Data and representations Expanded search Core	Abstract/project	Optical tactile images	Self-supervised tactile encoders with a multi-task TacBench evaluation suite.	Useful optical encoder initialization; frame/sample counts should not be presented as robot trajectories.
P45 · Sparsh-X 2025/06/17 Full citation	Data and representations Expanded search Supporting	Abstract/project	DIGIT 360 image, audio, motion, pressure	Learns multisensory tactile representations and studies manipulation transfer.	Requires a richer sensor suite than FlexiTac; do not assume all pretraining inputs exist on the target robot.
P46 · UniTouch 2024/01/31 Full citation	Data and representations Expanded search Supporting	Abstract/project	Optical touch aligned with multimodal embeddings	Aligns heterogeneous tactile sensors to visual representations connected with language and sound.	Perception and semantic alignment reference, not an autonomous action-data generator.
P47 · AnyTouch 2 2026/02/10 Full citation	Data and representations Expanded search Supporting	Abstract/project	Dynamic optical tactile and force data	Studies general dynamic tactile representations and introduces ToucHD.	Useful dynamics-oriented encoder benchmark; verify task-specific data access and sensor coverage.
P48 · Octopi 2024/05/05 Full citation	Data and representations Expanded search Supporting	Abstract/project	Tactile video + language / object properties	Reasons about object properties with tactile-language models and a tactile video dataset.	Useful for semantic grounding; property annotations do not supply robot action trajectories.
P49 · Touch and Go 2022/11/22 Full citation	Data and representations Expanded search Supporting	Abstract/project	Human-collected visual and tactile observations	Provides natural paired vision-touch data for multisensory learning.	Perception pretraining source; not a robot demonstration dataset with calibrated control actions.
P50 · RoboGen 2023/11/02 Full citation	Agents and generation Expanded search Core	Methods/results	Task, scene and supervision programs	Automates robotic skill learning through generative simulation.	Direct conceptual prior art; adding a tactile channel alone needs a stronger empirical contribution.
P51 · GenSim 2023/10/02 Full citation	Agents and generation Expanded search Core	Abstract/project	Executable simulation tasks and demonstrations	Uses coding LLMs to generate task curricula and expert demonstrations.	Task-level diversity is already an explicit objective; compare contact-conditioned generation beyond generic task variation.
P52 · GenSim2 2024/10/04 Full citation	Agents and generation Expanded search Core	Abstract/project	Articulated simulation + planning/RL teachers	Generates more realistic and articulated tasks with multimodal reasoning and policy training.	Strong environment-and-teacher baseline; tactile simulation and calibration remain additional work.
P53 · RoboTwin 2.0 2025/06/22 Full citation	Agents and generation Expanded search Core	Abstract/project	Bimanual simulation and execution code	Combines generated expert code with simulation feedback and structured domain randomization.	Mandatory baseline for code-generated data; existing visual data scaling is not evidence of tactile necessity.
P54 · MimicGen 2023/10/26 Full citation	Agents and generation Expanded search Core	Abstract/project	Transformed robot demonstrations	Scales human demonstrations through object-centric transformations in simulation.	Use as a fixed data-generation baseline; transformed trajectories still need contact validity checks.
P55 · ChronoAgentic (formerly Coding Agent Is Good As World Simulator) 2026/05/14 Full citation	Agents and generation Expanded search Core	Methods/results	Executable PyChrono scenes + visual and physics review	Plans, writes, executes and repairs physics simulation code using multiple reviewers.	Latest v3 evaluates 80 physics demos; video judging has shared-model and clip-length confounds and does not validate calibrated tactile policy transfer.
P56 · Code World Model 2026/08/26 Full citation	Agents and generation Expanded search Supporting	Methods/results	Persistent code state + proxy/video rendering	Separates executable world-state evolution from learned visual generation.	Architectural reference; generated visual observations are not calibrated tactile labels or contact-force validation.

Limitations and unresolved questions

This is a targeted literature map, not a systematic review of every tactile paper. The 56 entries are selected for sensor simulation, contact modeling, policy learning, data generation, or coding agents. Supporting perception and backbone papers are marked separately so that the count does not imply 56 equally direct competitors.

Methods/results coverage means the relevant sections were inspected. Abstract/project coverage is adequate for discovery and broad categorization but not for selecting a final reproduction recipe. Repository availability does not imply a complete release, license compatibility, reproducibility, or that installation was tested. No simulator, learned model, or physical robot experiment was executed for this review.

The most material remaining uncertainties are pressure-array calibration fidelity, sensor-specific transfer of pretrained encoders, access to complete action-labeled tactile datasets, teacher reliability under generated contacts, and the amount of real data needed for useful transfer. These should drive implementation decisions instead of an unsupported claim that the full pipeline is already solved.