Concept disambiguation

VLA vs VDA vs world models vs HFM

Compare model roles from the point of view of a VLA system: what enters the model, what comes out, and which part of the robot’s decision loop it can support. These labels describe overlapping dimensions.

By RoboSkin.ai Editorial Team · Source review: October 14, 2026

Expand VDA before comparing architectures

A unique, authoritative expansion of VDA was not found in the reviewed primary sources. This page uses “vision-conditioned diffusion action policy” as an explicit working definition, sometimes informally expanded as “Vision-Diffusion-Action”. The cited authors call their methods Diffusion Policy and Robotics Diffusion Transformer. The shorthand should not be read as their own taxonomy. If a source means video diffusion or dynamics by the same letters, inspect its actual prediction target before applying this comparison.

Compare definitions, inputs and outputs

Scroll horizontally to compare all fields. Sources in each row support its technical description and examples.

VLA, VDA, world model and HFM comparison with definitions, inputs, outputs, examples, relationships and sources.
Concept	Definition	Inputs	Outputs	Representative models	Relationship to VLA	Primary sources
VLAVision-language-action	An observation-and-instruction-to-action interface: visual context and language condition robot action generation.	Camera observations and language; robot state and contact sensing depend on the implementation.	Action tokens, continuous motor commands or action chunks.	RT-2, OpenVLA, RDT and π0; VLA-Touch adds a modular tactile feedback path.	Describes what the model connects. It does not require one action-generation algorithm or a predictive world model.	RT-2: the VLA formulation ↗OpenVLA ↗
VDAWorking definition: vision-conditioned diffusion action policy	An ambiguous shorthand. Here it means generating robot actions through a diffusion process conditioned on visual observations; it is not treated as a standardized model class.	Visual observations and usually robot state; some implementations also condition on language.	A denoised action or action sequence. Generating an action does not itself predict the next observation.	Diffusion Policy illustrates the mechanism without language; RDT is a language-conditioned diffusion policy. These papers do not need the VDA label.	Overlaps with VLA when language also conditions actions. A diffusion action head alone is not a world model.	Diffusion Policy: action diffusion ↗RDT: language-conditioned diffusion ↗
World modelPredictive environment dynamics	A model of how an environment changes, often conditioned on actions. Its predictive representation may be latent rather than a rendered video.	Observations or latent state, with action conditioning where the model supports controllable rollouts.	Predicted future state, observations, rewards or related dynamics quantities.	DreamerV3 learns environment dynamics and improves a separate behavior policy through imagined futures.	Can support training, planning or evaluation for a policy. Prediction and motor control remain distinct interfaces even inside a combined system.	DreamerV3: learning through world models ↗
HFMHaptic foundation model	A proposed reusable foundation for heterogeneous contact signals; the scope overlaps with tactile foundation models.	Touch / contact observations; the cited proposal discusses multimodal alignment and action-coupled sensing.	Reusable haptic representations, with action-conditioned prediction and active sensing among the proposed objectives.	Sparsh and UniTouch illustrate reusable touch representations. The HFM proposal is not a new universal checkpoint.	Could supply contact features to a VLA or tactile world model. A representation benchmark does not establish a closed-loop robot policy.	The Potential of Haptic Foundation Models ↗Sparsh ↗

VLA names an interface, not one architecture

Start with the output contract: does the system produce executable actions, or text that another controller must interpret? RT-2 puts actions into a token representation, while π0 uses flow matching for continuous action generation. Both connect visual context and language to action. PaLM-E is relevant to the history but produces high-level text / decisions, with a separate execution policy.

Use the existing VLA glossary definition for the short form, then inspect the model’s exact inputs and output role.

Diffusion describes how an action is generated

Original Diffusion Policy learns a conditional denoising process over actions. It does not require language conditioning. RDT combines diffusion action generation with language instructions, so a model can fit both the VLA interface and the working VDA description used here.

Flow matching, used by π0, is a related generative construction with its own training objective; it should not silently be relabeled as the original Diffusion Policy algorithm. Neither an action diffusion model nor a flow-matching action head automatically supplies a model of future observations.

A world model predicts consequences

A policy asks which action to take; a predictive dynamics model asks how the environment might evolve. DreamerV3 learns environment dynamics and improves behavior through imagined futures. The learned world model and the behavior policy serve different roles within that system.

For contact-rich robotics, the prediction target may be a tactile image, force-related state or latent contact representation. Use the existing tactile world model definition and world-model evidence guide to check action conditioning, rollout quality and whether prediction improves robot control.

HFM concerns reusable contact knowledge

From a VLA integration perspective, a reusable haptic encoder could supply contact features to an action policy; a haptic dynamics model could instead predict consequences for planning. These are possible integration roles, not capabilities established for every HFM. The HFM proposal distinguishes its broader design objectives from the frozen-encoder tasks actually benchmarked.

The detailed terminology discussion already lives in tactile foundation model vs haptic foundation model. Follow that comparison for the scope and evidence limits, or the HFM glossary entry for the concise definition.

Read a system by its interfaces

Check what observations are available at inference, what is predicted, which controller executes the result, and which experiment supports the claimed behavior. In VLA-Touch, touch contributes to high-level planning and action refinement around a base policy. That modular composition shows why a tactile encoder, VLA and controller can coexist without becoming the same model category.

Compare the existing tactile integration mechanisms →

Evidence and corrections

This is an editorial reading aid based on the linked primary sources. Results are author-reported; RoboSkin has not reproduced these model evaluations. Artifact availability and license terms can change.

Research methodology · Suggest a source or correction