Executive Summary

Key Figures

PRISM's story begins with these four numbers landing in one place. The input shrinks to a single photo, the human operator disappears, and yet performance goes up rather than down.

Source: PRISM (arXiv:2607.04880)

The recent way to make a robot handle objects on its own is the VLA (Vision-Language-Action) model. It takes the scene the camera sees and the instruction a person gives ("put the milk in the basket"), and directly outputs the arm's next move. Such policies learn by watching and imitating human demonstrations. So success in training comes down to "how much good demonstration data can you secure, and how good is it."

For a long time there were two roads to producing that data. The first is teleoperation. A person moves the robot directly with a joystick or a master arm, recording one successful demonstration at a time. You get high-quality data that matches the target environment exactly, but every sample passes through a human hand, so it is expensive, slow, and hard to scale.

The second is simulation. Inside a virtual environment you move the robot thousands of times and mass-produce demos. It is cheap and fast, but it has a catch. The scenes inside a simulator do not resemble a real user's kitchen, warehouse, or workbench. Lighting, objects, and layout all differ. Move a policy trained that way into reality and performance collapses, the so-called sim-to-real gap.

Figure 1. The two paths to robot demonstration data, and where PRISM sits. Teleoperation gives target-environment alignment; simulation gives scale and diversity — neither gave both. Pebblous original diagram (paper concept, reinterpreted).

PRISM's input is simple. The user takes one shot of their own workspace with an RGB-D camera (a camera that also captures depth) and gives one instruction like "put the milk in the basket." From there the pipeline goes to work. First a vision-language model recognizes what the objects in the photo are, and Grounded-SAM cuts out each object's boundary. The photo now becomes a structured scene: "one basket, one milk carton, and their positions."

The next step is the study's core idea, the digital cousin. Here the system does not copy the objects in the photo exactly. Instead it searches a large 3D asset library with CLIP and DINOv2 embeddings and pulls in different instances that are similar in meaning and appearance. It brings not the milk carton in the photo but that carton's cousins. Category and spatial layout follow the original photo, while each individual object is sampled anew every time.

Why a cousin, of all things? If you train only on a 'digital twin' that copies the original exactly, the policy overfits to that specific object and fails when color or shape changes even slightly. ACDC, the 2024 CoRL paper that first introduced the concept, reported that a policy trained on digital cousins was far more robust on a real robot than one trained on digital twins (90% vs. 25% zero-shot success). Appropriate diversity beats exact replication for generalization. PRISM takes this insight as the foundation for scene generation.

Figure 2. How one photo becomes a digital cousin scene. Instead of cloning the original, PRISM keeps the category and spatial layout and retrieves a different, similar instance via CLIP/DINOv2 embeddings. Pebblous original diagram (paper concept, reinterpreted).

Prior work had already reached the point of building digital-cousin scenes. Where PRISM goes one step further is that it also produces executable demonstrations on top of those scenes, without a human. With only a scene, the robot still doesn't know "so how am I supposed to move." Two mechanisms fill in that trajectory.

3.1Motion-Aware Grasp Selection

How the robot grips an object governs the entire motion that follows. Grip at an awkward angle and the arm has to take a contorted path, and such trajectories are hard to imitate. PRISM chooses the grasp that requires the least rotation from the robot hand's default orientation. The result is a trajectory where the arm moves naturally, one that favors imitation learning.

3.2Keep the Trajectory, Shake Only the Surface

Once the grasp is set, a physically valid trajectory is generated procedurally inside the simulator. A human joystick never appears once. The last mechanism that attaches data diversity is trajectory-preserving randomization. The path of the motion itself stays fixed, while lighting, texture, camera angle, and the object's initial pose are varied to film the same action hundreds of times differently. So the policy does not depend on a particular light or background: keep the essence (the trajectory), shake only the surface (the appearance).

Figure 3. The two mechanisms behind demonstrations with no human in the loop. PRISM picks the grasp with the least rotation for a natural trajectory (①), then fixes that trajectory and varies only lighting, texture, and camera angle (②, dashed orange = the same trajectory). Pebblous original diagram (paper concept, reinterpreted).

The effect was measured on two stages. One is the LIBERO and LIBERO-Plus generalization benchmarks that compare simulation against simulation; the other is real-world experiments using an actual robot. The table below shows PRISM's success rate against two baselines (RoboTwin 2.0, X-Sim) on two manipulation tasks. The columns to watch are LIBERO and LIBERO-Plus, which handle variations unseen during training.

Unit: success rate (%). Source: PRISM (arXiv:2607.04880), Table. In-Domain is within the training distribution; LIBERO and LIBERO-Plus measure generalization to unseen variations.

The pattern is clear. In-Domain, the three methods trade places, but move to LIBERO and LIBERO-Plus, which handle variations unseen during training, and PRISM opens a large gap. On the first task's LIBERO it scores 98.0%, leading X-Sim (48.0%) by 50 percentage points, and on LIBERO-Plus it reaches 67.6%, nearly double. It means data that secures both target-environment alignment and diversity pays off in generalization.

The real-robot experiments pointed the same way. On three real-world manipulation tasks, putting milk in a basket, placing a wine bottle in a cabinet, and putting a box in a basket, policies trained with PRISM reached up to 100% success and held their robustness under environment variations different from training. The weight of these numbers is that they come from data no human ever teleoperated even once.

The authors also state the limits plainly. Today's PRISM is confined to the Franka arm and rigid objects. It cannot yet handle deformable objects like cloth, rope, or food. And because it reconstructs 3D from a single image, when an object is heavily occluded the reconstruction becomes inaccurate, and demonstration quality can drop accordingly. The strength of a single photo is also, at once, the constraint of a single photo.

This is not to claim PRISM alone has solved the robot-data problem. The study is closer to one case within a larger current. Around the same time, work such as Real2Render2Real and Robot Learning from Any Images is growing along several branches, sharing the problem statement of "make robot learning data out of images of the real site." The basic verb for securing robot data is shifting from "collect" to "generate."

This shift returns a familiar question to anyone who works with data. In the era of scraping data off the web, the crux was "how much, and how diverse, did you collect." Once you start making data to fit the site, the crux moves to "does the data you made truly represent the real site, and is its diversity sufficient and unbiased?" The easier generation gets, the more important verification becomes. Because mass-producing data that looks plausible but is misaligned with the real site gets just as easy.

Source Paper

• 1.Ko, D., Kim, H., Yeo, C., Lee, D., Park, T., & Hwang, H. (2026). "PRISM: Personalized Robotic Dataset Generation via Image-based Scene and Motion Synthesis." arXiv:2607.04880 (Kyung Hee University).

Lineage & Adjacent Work

• 2.Dai, T., et al. (2024). "ACDC: Automated Creation of Digital Cousins for Robust Policy Learning." CoRL 2024, arXiv:2410.07408.
• 3.Yu, J., et al. (2025). "Real2Render2Real: Scaling Robot Data Without Dynamics Simulation or Robot Hardware." arXiv:2505.09601.
• 4.Authors et al. (2025). "Robot Learning from Any Images." arXiv:2509.22970.