Multimodal AI

Summary

Multimodal AI builds systems that take in more than one kind of signal at once, such as images, language, audio, depth, force, and motion, and produce a single interpretation or a single action from them. It matters because the physical world does not arrive as one clean stream. A robot deciding whether it has actually grasped a cup needs vision and force together, and either signal alone gives a confident wrong answer. The central tension of the field is that the value lives in the relationship between signals rather than in any one of them, so the hard problem is not collecting more data but proving that every part of one sample describes the same moment. That makes most reliability work in multimodal AI alignment and evaluation work rather than model work.

Multimodal AI: Definition and Scope

Multimodal AI covers any system whose inputs, outputs, or internal representation span two or more modalities that carry genuinely different information. The standard academic decomposition comes from Baltrusaitis, Ahuja and Morency, whose 2017 survey Multimodal Machine Learning: A Survey and Taxonomy names five core challenges: representation, translation, alignment, fusion, and co-learning. Those five still describe the field accurately, and alignment is the one that industry consistently underestimates.

In scope: shared embedding spaces that place image and text in one geometry, Sensor Fusion that combines LiDAR, radar and camera into one scene estimate, Vision-Language-Action Models that emit robot actions from pixels and an instruction, and generative models that produce video, audio and control signals together.

Multimodality becomes operationally important in Physical AI, where a machine must connect language and perception with sensor state and physical action. It is equally foundational to Embodied AI, where an agent learns through the closed loop between what it perceives, what it does, and how the environment changes. These fields overlap, but they are not interchangeable: multimodal AI describes how different signals are represented and combined, while physical and embodied AI describe intelligence acting through a body in an environment.

Explicitly out of scope: a system that merely accepts several file types and routes each to a separate single-modality model. If the modalities never meet inside a shared representation, the pipeline is a set of unimodal systems behind one interface. A language model also does not become multimodal because captions appeared somewhere in its training corpus. The test is whether the model’s output changes when one modality contradicts another.

The industry term and the academic term diverge in one useful way. Research usually says multimodal when it means learned joint representation. Robotics and autonomy teams often say multimodal when they mean multi-sensor capture, which is an upstream data property rather than a model property. Both usages appear in this graph, and the distinction matters because a multi-sensor dataset can be recorded correctly and still be useless for joint representation learning if its streams are not aligned in time.

How the Field Arrived Here

The pattern the table reveals is a steady widening of what counts as a modality. Text and image were joined first, then audio and depth, then action itself. Action becoming a modality that models generate rather than a separate control stack is the shift that pulled multimodal AI into robotics.

Why Now

Representation alignment became a reusable primitive. Before CLIP, joining two modalities often meant designing a fusion architecture for each task. After it, teams could start from a pretrained shared embedding space rather than learning that relationship from scratch. This lowered the model-design cost of adding scale and helped general multimodal backbones emerge, even though aligning the underlying data across modalities remained a separate operational problem.

One architecture now absorbs every modality. Tokenization plus attention removed the need for modality-specific encoders as a design commitment. Cosmos 3 is the current end point of that trend, described by NVIDIA as an omnimodel that natively understands and generates text, images, video, ambient sound and actions in one system.

Cross-embodiment pooling proved that transfer is real. Open X-Embodiment was the first large test of whether data from one robot helps another, and RT-1-X improved on robot-specific methods by roughly 50 percent on small datasets. That result changed data strategy across the field, because it made a shared corpus worth contributing to.

Synchronized capture hardware got cheap enough to deploy at scale. Ego-Exo4D captured more than 800 participants across 13 cities with time-aligned egocentric and exocentric rigs. Multi-sensor capture stopped being a lab exercise and became something an operations team can run.

Generation closed the loop back to control. World models that emit joint angles, gripper positions and trajectory points rather than only video mean synthetic data can be used for policy training and evaluation, not only for perception pretraining.

None of these solved alignment at the data layer. Every one of them assumes that when a sample arrives with a camera frame, a force reading and an instruction attached, those three things genuinely belong to the same instant. Nothing in the model stack verifies that assumption, and nothing in it recovers when the assumption is false.

Multimodal AI Compared to Single-Modality AI

The time reference row causes the others. Once correctness depends on a shared clock, the unit of data becomes a bundle rather than a record, missing inputs become indistinguishable from stale ones, evaluation has to probe disagreement rather than accuracy, and a repair reaches back through an entire session instead of a set of rows. Teams that treat timestamps as metadata inherit all six of the other rows as surprises.

Why Multimodal AI Matters

The practical challenge is that multimodal data can be wrong in a way that passes every check a unimodal pipeline knows how to run. Each stream is well formed, the file counts match, the annotations are complete, and the dataset is still corrupt because the streams describe different moments.

Take one concrete task: a mobile manipulator in a warehouse picking a deformable item, a shrink-wrapped pack of paper towels, out of a mixed tote. To learn this from demonstration the system needs an RGB frame at the moment of contact, a depth estimate of the tote’s contents, force and torque readings through the closing motion, joint states for the arm, the language instruction that scoped the task, and a success signal recorded after the item cleared the tote. Six streams, at different native rates. The camera runs at 30 Hz, so a frame arrives every 33 milliseconds. The force torque channel runs near 1 kHz. The label that matters, the instant of first contact, is a relationship between two of those streams and exists in neither one alone.

Now let one clock drift. If the force channel’s timestamps slip by 120 milliseconds across an eight hour collection shift, first contact gets attributed to a frame roughly four camera frames before the gripper actually touched anything. Every affected sample teaches the policy to begin closing early. Nothing in the dataset looks wrong: the images are sharp, the forces are plausible, the annotations are complete, and validation accuracy is fine because the validation set drifted identically. The failure surfaces in production as crushed packs, and by then the training corpus has thousands of samples carrying the same bias, with no field recording which of them are affected. A correct label on the wrong moment is still incorrect data, and it is the most expensive kind because it is invisible to every quality check that inspects one modality at a time.

That is the failure mode worth designing around, and it is a data operations failure rather than a modeling one. No amount of architecture search recovers a corpus whose time reference was never established.

Industry Perspective

Leading organizations disagree materially about architecture. Physical Intelligence trains a single end-to-end policy and argues that co-training on heterogeneous sources is what produces open-world generalization. Google DeepMind splits the problem, putting embodied reasoning in one model and motor control in another, and reports skills transferring between an ALOHA2 platform, a Franka arm and Apptronik’s Apollo humanoid. NVIDIA bets that a generative world model can supply much of the training and evaluation data, positioning Cosmos 3 as a way to compress physical AI cycles from months to days. Meta invests in human capture rather than robot capture, on the view that egocentric human video is the scalable source. These are not compatible strategies, and at least one of them will look wrong in three years.

They converge on three things anyway. All four publish or consume corpora where a sample is a bundle of time-aligned streams rather than a file. All four treat action as a modality inside the model rather than as a downstream controller. And all four have made cross-embodiment transfer, moving data or skills between robot bodies, an explicit design goal rather than an afterthought.

The inference that convergence supports is that the durable asset in this field is the sample contract, not the model. Model families and architectures have changed rapidly since 2021, while many of the same underlying corpora have remained useful across successive generations. Whichever architecture wins consumes samples whose modalities are aligned to a stated reference and whose gaps are declared. A team that gets the contract right can retrain on a new architecture in weeks. A team that got it wrong recollects.

Kotwel’s Role

Kotwel does not build robots or train foundation policies. That boundary is deliberate because it keeps Kotwel complementary to robotics programs regardless of the hardware, model family, or control architecture they use. Kotwel works on the data operation underneath those systems.

For multimodal programs, that role begins with defining the sample contract: the reference clock shared by all streams, the acceptable alignment tolerance for each modality, and the states used to record missing or unusable inputs. These decisions turn a collection of camera frames, sensor readings, instructions, and actions into traceable samples that can be validated consistently.

The operational pipeline then enforces that contract. Ingestion normalizes streams against the declared time reference and checks timestamp consistency. Annotation records events against that reference rather than against the index of one stream. Cross-modal review checks whether the selected instant is consistent across the relevant modalities, while release validation identifies drift, undeclared gaps, and other session-level failures before the data enters training.

Kotwel’s scope ends with a verified dataset and its supporting provenance, schema, and quality records. Kotwel does not evaluate the customer’s policy or promise a model outcome. Its contribution is to make the multimodal evidence entering that model synchronized, traceable, and auditable.

Key Concepts

• Sensor Fusion combines readings from several sensors into one estimate of the world’s state, usually with an explicit uncertainty model.

• Cross-modal Alignment establishes which parts of two streams refer to the same event, and is the challenge that most multimodal failures trace back to.

• Data Synchronization is the operational practice of putting every stream on a shared reference clock and proving the offsets stay inside tolerance.

• Multimodal Learning covers the training methods that build a joint representation rather than routing modalities to separate models.

• Contrastive Pretraining learns a shared embedding by pulling matched pairs together and pushing mismatched pairs apart, the method behind CLIP.

• Vision-Language-Action Models emit robot actions directly from visual input and a language instruction, treating action as an output modality.

• World Models learn to predict how a scene evolves, and increasingly emit control outputs alongside generated video.

• Missing Modality Robustness concerns whether a model degrades gracefully when a stream drops out, which requires deliberate dropout testing to measure.

• Egocentric Data captures the world from a person’s own viewpoint and is the main scalable substitute for robot-collected demonstrations.

• Teleoperation Data records a human driving a robot directly, producing the aligned observation and action pairs that imitation learning needs.

• Tactile Sensing supplies contact information that vision cannot recover, and is the modality most often omitted and later missed.

• Synthetic Data generated from world models supplies coverage for rare events, at the cost of a domain gap that has to be measured rather than assumed.

• Multimodal Evaluation tests behavior under contradiction and dropout, because held-out accuracy on an identically assembled test set cannot detect assembly errors.

• Building Reliable Multimodal AI Training Datasets covers the sample-level construction practices this hub summarizes.

Further Reading

Primary sources

• Multimodal Machine Learning: A Survey and Taxonomy, Baltrusaitis, Ahuja and Morency, 2017.

• Open X-Embodiment: Robotic Learning Datasets and RT-X Models, Open X-Embodiment Collaboration, 2023.

• Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives, Grauman et al., 2023.

• pi-0.5: a Vision-Language-Action Model with Open-World Generalization, Physical Intelligence, 2025.

• Gemini Robotics 1.5, Google DeepMind, 2025.

Industry and strategy

• Cosmos 3 launch announcement, NVIDIA, May 2026.

• Gemini Robotics brings AI into the physical world, Google DeepMind, March 2025.

• Ego-Exo4D documentation and dataset cards, Meta AI.