In computer vision and natural language processing, models generalize because their inputs share a universal structure: pixels are pixels, and text tokens are text tokens. Robotics, however, has historically been trapped in hardware silos. A control policy trained on a specific seven-degree-of-freedom robot arm cannot inherently control a six-degree-of-freedom arm, let alone a bimanual humanoid or a mobile manipulator.
This friction arises from the embodiment gap. Traditional imitation and reinforcement learning frameworks map visual observations directly to platform-specific motor commands (such as joint angles or torques). If the physical structure of the robot changes, or if its camera is shifted by a few centimeters, this direct mapping breaks. Furthermore, because a robot's own chassis or gripper often appears in its camera feed, policies frequently condition on the specific visual appearance of their training hardware. As noted by Alexander Khazatsky et al. in their introduction of the DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset, restricting data to a single hardware platform prevents models from developing the physical flexibility required for true generalist intelligence.
To build universal robotic foundation models, researchers are developing techniques to decouple a policy's core physical reasoning from the specific hardware executing the task.
Decoupling Kinematics via Shared Action Spaces
The most direct way to bridge the embodiment gap is to eliminate joint-space control entirely. If two robots have completely different internal joint configurations but operate in the same three-dimensional world, their actions can be translated into a shared coordinate system.
One intuitive approach is to define actions within the operational or end-effector space. As demonstrated by Roberto Martín-Martín et al. in Variable Impedance Control in End-Effector Space: An Action Space for Reinforcement Learning in Contact-Rich Tasks, abstracting control to end-effector trajectories provides a buffer against underlying kinematic differences, allowing policies to achieve zero-shot transfer across distinct mechanical structures.
When physical forms diverge radically—such as transferring skills between a parallel-jaw gripper and a five-fingered dexterous hand—hand-engineered coordinate spaces are insufficient. To solve this, researchers use contrastive learning to construct latent action spaces. For example, Elvis Nava et al. developed a framework in Latent Action Diffusion for Cross-Embodiment Manipulation that maps heterogeneous control signals into a unified latent space. By training a diffusion policy within this shared space, a single model learns to guide multiple robot types, allowing data collected on simple grippers to directly improve the performance of complex dexterous hands. By aligning diverse platforms into a shared egocentric action space, as explored by Chelsea Finn in Pushing the Limits of Cross-Embodiment Learning for Manipulation and Navigation, policies can even transfer structural insights across entirely different domains, such as using mobile navigation data to improve stationary manipulation.
Shifting Focus to Object-Centric Dynamics
An alternative strategy for achieving embodiment invariance is to ignore the robot's physical structure entirely during high-level planning. Instead of asking "How should the robot move its arm?", the policy asks "How should the target object change?" Once the desired environmental change is determined, a low-level, platform-specific controller handles the precise inverse kinematics required to execute it.
This concept underpins object-centric and task-space planning. Weiliang Tang et al. formalize this in Embodiment-Agnostic Action Planning via Object-Part Scene Flow by predicting 3D scene flow at the object-part level. Because the model's output is strictly a prediction of how an object's components should move in space, the resulting plan can be executed by any robot capable of generating the necessary forces.
Similarly, Yifeng Zhu et al. demonstrate in Learning Generalizable Manipulation Policies with Object-Centric 3D Representations that transforming 2D camera feeds into object-centric 3D point clouds strips away the visual presence of the robot arm. By anchoring the policy's attention entirely to the geometry of the object, the model becomes naturally invariant to changes in both camera viewpoints and robot morphology.
Visual Alignment and Morphological Invariance
Even when action spaces are unified, a perceptual embodiment gap remains. If a policy is trained on a robot with a blue parallel gripper, deploying it on a robot with a metallic dexterous hand introduces out-of-distribution visual features that can cause neural networks to fail.
To resolve this, researchers use visual translation and masking techniques to standardize the policy's visual inputs. The simplest approach is to remove distracting surface features entirely. Nawid Jamali et al. demonstrate in Robot Bed-Making: Deep Transfer Learning Using Depth Sensing of Deformable Fabric that training models exclusively on depth images provides robust invariance to color, texture, and minor structural variations across different robot arms.
More advanced techniques actively alter the visual stream to preserve a consistent input distribution. Marion Lepert et al. introduce this in Shadow: Leveraging Segmentation Masks for Cross-Embodiment Policy Transfer by overlaying composite segmentation masks onto the robot's end-effector, neutralizing visual differences between the source and target hardware.
Taking this concept further, Lawrence R. Chen et al. develop a visual "cross-painting" mechanism in MIRAGE: Cross-Embodiment Zero-Shot Policy Transfer with Cross-Painting. Mirage dynamically tracks the target robot's arm in the live video feed, masks it out, and inpaints a 3D rendering of the original source robot used during training. This visual substitution tricks a hardware-specific policy into operating successfully on an entirely new physical platform without requiring any retraining.
Modular Architectures: Stems and Trunks
To ingest data from a wide variety of robots—each with different numbers of joints, sensor arrays, and sampling rates—neural network architectures must move away from rigid, monolithic designs. Instead, the field is converging on modular architectures that separate hardware-specific processing from general physical reasoning.
This paradigm splits the network into embodiment-specific "stems" and a shared, universal "trunk." As detailed by Lirui Wang et al. in Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers, lightweight input stems are customized for each robot configuration to ingest unique proprioceptive data and project it into a standardized token space. These tokens are then processed by a massive, shared transformer trunk that models core physical interactions and task strategies.
[ Robot A Stem ] ---\
[ Robot B Stem ] ----+---> [ Shared Transformer Trunk ] ---> [ Hardware-Specific Action Heads ]
[ Robot C Stem ] ---/
This modularity allows researchers to pool data from hundreds of distinct robotic platforms into a single training pipeline, ensuring that the central model benefits from collective data scale. Similarly, Yifan Zhou et al. utilize hierarchical modularity in Learning modular language-conditioned robot policies through attention to construct reusable functional blocks, ensuring that language-conditioned policies remain stable when transferred across differing arm morphologies.
Scaling Beyond Silos
By combining shared action abstractions, object-centric planning, visual alignment, and modular neural architectures, cross-embodiment generalization provides a scalable path forward for robotic learning. Large-scale, multi-environment datasets like BridgeData V2 (Homer Rich Walke et al.) and DROID (Alexander Khazatsky et al.) provide the structural diversity needed to train these models. As these techniques mature, the process of deploying a policy to a new robot will shift from weeks of custom data collection to zero-shot, plug-and-play transfer.