Author: Denis Avetisyan
A new tri-manual robotic system demonstrates advanced manipulation skills by autonomously preparing sashimi from a salmon fillet.
This work presents Sashimi-Bot, a robotic platform integrating visual and tactile sensing with deep reinforcement learning for precise deformable object manipulation.
Despite advances in robotic manipulation, reliably handling deformable objects with inherent variability remains a significant challenge. This is addressed in ‘Sashimi-Bot: Autonomous Tri-manual Advanced Manipulation and Cutting of Deformable Objects’, which introduces a tri-manual robotic system capable of autonomously preparing sashimi from a salmon loin. By integrating deep reinforcement learning with visual and tactile feedback, the system demonstrates robust grasping, cutting, and slicing of this highly deformable food item. Could this work pave the way for more adaptable robots capable of tackling complex manipulation tasks in unstructured, real-world environments?
Deconstructing Culinary Tradition: The Automation Imperative
The foundation of global food systems currently rests on a substantial workforce dedicated to manual labor, a reality that introduces inherent limitations to both efficiency and expansion. From harvesting crops to the intricate processes of food preparation, human hands perform tasks demanding dexterity, judgment, and adaptability – skills difficult to replicate mechanically. This reliance creates predictable bottlenecks in the supply chain, hindering the ability to rapidly scale production to meet growing demands or respond effectively to disruptions. Consequently, food prices can fluctuate based on labor availability, and maintaining consistent quality across large volumes becomes a considerable challenge. Addressing this dependence through automation isn’t simply about replacing jobs; it’s about building a more resilient and sustainable food future capable of feeding a growing population with predictable output and consistent standards.
The automation of preparing delicate foods, such as sashimi, presents a unique set of robotic manipulation challenges far exceeding those encountered in typical industrial automation. Unlike assembling rigid parts, handling fish fillets requires discerning subtle textural differences to avoid crushing or tearing the protein, demanding sensors capable of ‘feeling’ for the ideal consistency. Existing robotic grippers often lack the dexterity and sensitivity to replicate the precise movements of a skilled chef, and vision systems must accurately identify and adapt to the inherent variability in natural food products – differing size, shape, and the presence of skin, bones, or fat. Successfully mimicking these skills necessitates advancements in soft robotics, tactile sensing, and adaptive control algorithms, allowing robots to not simply grasp but to feel and respond to the delicate nature of the food, potentially revolutionizing culinary practices and food production.
The potential for automated food preparation extends beyond mere convenience, offering pathways toward significantly more efficient and sustainable food systems. By precisely controlling portion sizes, minimizing waste through optimized cutting and ingredient use, and reducing the need for long-distance transportation of prepared foods, automation can address critical environmental concerns. Furthermore, automated systems can operate continuously, increasing overall food production capacity without requiring expanded agricultural land or increased labor demands. This shift allows for localized food production, reducing reliance on complex supply chains and enhancing food security, particularly in urban environments and regions with limited resources. Ultimately, automating delicate culinary tasks like sashimi preparation serves as a microcosm of a broader opportunity – reshaping how food is produced, distributed, and consumed for a more resilient future.
Sashimi-Bot: A Tri-Manual System for Controlled Dissection
Sashimi-Bot is a robotic system designed to automate the process of preparing sashimi from salmon loins. The system utilizes three 7-DoF robotic arms to execute the necessary manipulation, cutting, and placement tasks. These 7-DoF arms provide the dexterity required for complex movements within the constrained workspace, enabling precise control during each stage of the preparation process. The robotic configuration allows for a high degree of freedom, facilitating the manipulation of the salmon loin and the execution of precise cuts required for traditional sashimi presentation.
Sashimi-Bot achieves fully autonomous sashimi preparation through a sequential three-stage process. Initially, shape manipulation adjusts the salmon loin to a suitable configuration for cutting. This is followed by the cutting stage, where precise slices are created using the robotic arms. Finally, the picking & placing stage utilizes the robots to transfer the prepared sashimi portions to a designated output location, completing the automated process. Each stage is interdependent and contributes to the overall functionality of the system, enabling complete, unattended operation.
Sashimi-Bot utilizes an RGB-D camera and a GelSight sensor to enable precise robotic manipulation and cutting of salmon. The RGB-D camera provides color and depth information, allowing the system to perceive the 3D structure of the salmon loin and locate relevant features for processing. Complementing this, the GelSight sensor—a tactile sensor—provides high-resolution, textural feedback during manipulation and cutting. This tactile data is crucial for confirming contact, assessing cutting force, and detecting slippage, which are essential for consistent and accurate sashimi preparation. The combined data streams from both sensors facilitate closed-loop control, allowing the robot to adapt to variations in salmon texture and shape and maintain the required precision throughout the automated process.
Deformable Object Manipulation: A Study in Material Response
The salmon loin utilized in this research exhibits elastoplastic material properties, meaning it undergoes both elastic deformation – temporarily changing shape under force and returning to its original form – and plastic deformation – undergoing permanent shape changes. This combination presents a manipulation challenge because the material’s response is non-linear and history-dependent; prior deformations influence its subsequent behavior. Specifically, the salmon’s composition, including muscle fibers and connective tissues, results in anisotropic deformation characteristics – differing responses to force based on direction – and a susceptibility to both bending and stretching under relatively low forces, necessitating precise control during manipulation to avoid tearing or unwanted shape changes.
Sashimi-Bot employs non-prehensile manipulation, meaning it reshapes the salmon loin without grasping or encircling the object. This approach utilizes controlled force application to deform the material directly. Testing demonstrated a 100% success rate in manipulating the salmon loin from initial configurations resembling the letters L, C, and Z into a flat, fileted state. This consistent performance indicates the robustness of the non-prehensile technique for this specific elastoplastic material and these target geometries, avoiding the complexities associated with traditional grasping methods.
Shape manipulation of the salmon loin using a Deep Reinforcement Learning (DRL) approach demonstrated varying levels of complexity reflected in the number of actions required for successful deformation. Specifically, initial configurations shaped as an ‘L’ required an average of 1.8 actions to manipulate into the target configuration. More complex ‘C’ shaped configurations necessitated an average of 5.0 actions, while the most complex ‘Z’ shaped initial configurations required an average of 7.5 actions to achieve the desired shape. These action counts represent the number of discrete control steps taken by the robot to complete the manipulation task, indicating a direct correlation between geometric complexity and the computational effort needed for successful DRL-based shape manipulation.
The system employs a motion planning algorithm that utilizes data from a GelSight sensor to achieve precise and repeatable cutting paths. This sensor provides high-resolution tactile feedback, allowing the algorithm to dynamically adjust the cutting trajectory based on real-time contact information. The GelSight sensor detects subtle variations in pressure and surface geometry, enabling the system to compensate for inconsistencies in the salmon loin’s material properties and initial configuration. This feedback loop ensures consistent cutting performance across different attempts and object shapes, resulting in accurate and uniform sashimi preparation.
Precision and Efficiency: Validating the Automated Dissection
The Sashimi-Bot achieves precise manipulation of delicate sashimi slices through a sophisticated integration of visual perception and visual servoing techniques. By analyzing visual features – shape, size, and texture – the system identifies and localizes each slice with high accuracy. This visual data then drives a closed-loop control system, known as visual servoing, which guides the robotic arm to approach, grasp, and place the sashimi with remarkable precision. The robot doesn’t simply follow pre-programmed movements; it continuously adjusts its actions based on real-time visual feedback, allowing it to compensate for slight variations in slice position or orientation. This dynamic adaptation is crucial for handling the fragile nature of sashimi and ensuring consistent, accurate placement – a key component in both presentation and efficient food preparation.
The robotic system demonstrated a high degree of precision in both the cutting and retrieval of sashimi slices. During trials, the automated process successfully cut 28 out of 34 salmon slices to the desired specifications, indicating a robust capability in handling delicate food items. Following the cutting phase, the system reliably picked up 26 out of 28 slices directly from the cutting board, showcasing its proficiency in visual identification and manipulation of objects within a complex environment. These results suggest a significant step towards fully automated sashimi preparation, potentially offering increased throughput and consistent quality compared to traditional methods.
A particularly noteworthy achievement of the automated system lies in its ability to retrieve delicate sashimi slices directly from the knife blade. During testing, the robot successfully picked all six slices presented on the blade, demonstrating a precise and controlled manipulation capability. This feat is especially challenging due to the inherent slipperiness of the fish and the sharp, confined space, but the integration of visual servoing and tactile sensing allows for gentle yet secure grasping. Such precision isn’t merely a technical accomplishment; it represents a significant step towards fully automating the complex task of sashimi preparation and minimizing potential damage to the food during handling.
The system’s ability to reliably identify different sashimi slices hinges on a highly accurate tactile sensor classification model. This model, rigorously tested on a separate validation dataset, achieved a remarkable 95% accuracy in distinguishing between slices based on their textural properties. This level of precision is crucial for the robotic system to not only locate slices, but also to assess their quality and suitability for serving – effectively mimicking the skilled touch of a sushi chef. The high accuracy suggests the tactile data provides a robust and reliable input for the robot’s decision-making process, ensuring consistent and precise handling of each delicate piece of sashimi.
Sashimi-Bot represents a substantial leap forward in culinary automation, consistently delivering precisely cut and placed sashimi with a level of efficiency unattainable through manual methods. The system’s ability to autonomously manage the entire preparation sequence – from initial slicing to final placement – not only accelerates the process but also minimizes variations inherent in human execution. This consistent performance translates directly into reduced food waste, optimized resource allocation, and a standardized product quality, effectively addressing the challenges of maintaining high output and consistent results in professional kitchens. By removing reliance on skilled labor for repetitive tasks, Sashimi-Bot offers a scalable solution for restaurants and food processing facilities seeking to improve operational efficiency and maintain stringent quality control.
The automation of sashimi preparation, as demonstrated by this system, addresses a growing challenge within the culinary landscape: a persistent labor shortage. Restaurants and food service establishments increasingly struggle to find skilled workers capable of performing repetitive, detail-oriented tasks like slicing and plating delicate dishes. Beyond mitigating staffing concerns, this technology offers a significant boost to food safety protocols. By minimizing human contact with food products, the risk of contamination is substantially reduced, offering a more hygienic preparation process. This controlled environment, coupled with consistent precision, not only ensures a higher quality product but also supports adherence to increasingly stringent food safety regulations, promising a future where culinary excellence and public health go hand in hand.
The development of Sashimi-Bot exemplifies a willingness to challenge the established boundaries of robotic manipulation. The system doesn’t merely execute pre-programmed motions; it learns to interact with the inherent unpredictability of deformable objects like salmon. This echoes Brian Kernighan’s sentiment: “Debugging is like being the detective in a crime movie where you are also the murderer.” The ‘murderer’ in this case isn’t malicious, but the complexity of the task itself. Sashimi-Bot, through deep reinforcement learning and tactile sensing, systematically investigates and resolves the ‘crime’ of imperfect cuts and unstable manipulation, iteratively refining its approach to achieve mastery over a traditionally human skill. It’s a study in controlled chaos, extracting order from a flexible, yielding reality.
Beyond the Slice
The successful automation of sashimi preparation, while a charming demonstration, merely exposes the fragility of established assumptions regarding robotic manipulation. The system functions – it cuts fish. But what happens when the fish isn’t salmon? Or when the desired cut isn’t a simple slice, but a complex, artistically-motivated curve? Sashimi-Bot operates within carefully defined parameters; the true challenge lies in dismantling those parameters and forcing the system to generalize. The reliance on both visual and tactile feedback represents progress, yet the information still needs interpretation—and interpretation is inherently subjective. Can a robot truly ‘feel’ the resistance of the fish, or simply register a pressure value?
The path forward isn’t about refining the cutting algorithm, but about questioning the very notion of ‘skill’ in this context. Current approaches treat manipulation as a series of precisely controlled movements. What if the system embraced a degree of controlled chaos—allowing for slight deviations, learning from ‘mistakes’, and adapting in real-time to unexpected variations in the material? The creation of truly robust deformable object manipulation demands a willingness to relinquish control, to allow the robot to discover its own solutions, even if those solutions appear inefficient or unconventional.
Ultimately, Sashimi-Bot isn’t about making robots that can cut fish; it’s about understanding the limits of predictability. Each successful slice is a temporary reprieve, a localized victory against the inherent messiness of the physical world. The real breakthroughs will come when the system fails spectacularly, forcing a re-evaluation of the underlying principles and revealing the deeper, more subtle laws governing the interaction between robots and the objects they manipulate.
Original article: https://arxiv.org/pdf/2511.11223.pdf
Contact the author: https://www.linkedin.com/in/avetisyan/
See also:
- Clash Royale Best Boss Bandit Champion decks
- When Is Predator: Badlands’ Digital & Streaming Release Date?
- Mobile Legends November 2025 Leaks: Upcoming new heroes, skins, events and more
- eFootball 2026 Show Time National Teams Selection Contract Guide
- Clash Royale Furnace Evolution best decks guide
- You can’t watch Predator: Badlands on Disney+ yet – but here’s when to expect it
- Hazbin Hotel Season 2 Episode 5 & 6 Release Date, Time, Where to Watch
- Deneme Bonusu Veren Siteler – En Gvenilir Bahis Siteleri 2025.4338
- JoJo’s Bizarre Adventure: Ora Ora Overdrive unites iconic characters in a sim RPG, launching on mobile this fall
- Clash Royale November 2025: Events, Challenges, Tournaments, and Rewards
2025-11-17 15:58