{"slug": "sashimi-bot-uses-tactile-sensing-to-cut-salmon", "title": "Sashimi-Bot uses tactile sensing to cut salmon", "summary": "Researchers published Sashimi-Bot in npj Robotics, a three-armed robotic system that uses tactile sensing to cut salmon for sashimi. The system achieved 95% accuracy in detecting blade-to-board contact and successfully cut 34 slices, placing 26 of 28 that landed on the board. The work represents a milestone in robotic manipulation of deformable objects.", "body_md": "# Sashimi-Bot uses tactile sensing to cut salmon\n\nResearchers published Sashimi-Bot in npj Robotics, a three-armed robotic system that prepares sashimi from a raw salmon loin. Three manipulators divide roles: one arm holds and slices with a chef's knife fitted with a GelSight tactile sensor, a second stabilizes the fish, and a third picks up slices with chopsticks. Manipulation policies were trained via deep reinforcement learning in simulation and transferred to the real robot without additional real-world tuning, per the paper. The GelSight sensor, trained on 12,397 readings from 157 cutting motions, achieved 95% accuracy in detecting blade-to-board contact (99% precision, 67% recall), per ZME Science coverage. In full tests, the system cut 34 salmon slices and successfully placed 26 of 28 that landed on the board; six slices that stuck to the blade were all recovered.\n\n### What happened\n\nPer the npj Robotics paper (Herland et al., 2026), researchers developed Sashimi-Bot, a three-armed robotic system that prepares sashimi from a raw salmon loin. Three manipulators divide the task: one arm holds and slices with a chef's knife, a second stabilizes and shapes the salmon, and a third uses chopsticks to pick up and place individual slices, as described in the paper and in coverage by ZME Science and Interesting Engineering. The team trained manipulation policies through deep reinforcement learning in simulation and transferred controllers to the real robot without additional real-world tuning.\n\n### System details\n\nThe knife arm uses a GelSight tactile sensor - a soft gel surface with an embedded camera - to detect pressure and deformation at the blade tip, per ZME Science's coverage. The sensor was trained on 12,397 readings from 157 cutting motions. On a held-out test set, the model reached 95% accuracy in detecting board contact, with 99% precision but lower recall at 67%, per ZME Science. In full system tests, Sashimi-Bot cut 34 salmon slices; of the 28 that landed on the cutting board, the robot successfully picked up 26. The two failures involved very thin slices slipping from the chopsticks. Six slices stuck to the knife blade were all recovered by picking them directly from the blade, per ZME Science. A full cut cycle averaged 27.9 seconds, or 37.7 seconds when a blade recovery was needed.\n\n### Technical context\n\nThe paper combines sim-to-real transfer, multi-arm choreography, and contact-rich control for a class of objects - soft, slippery, and deformable - where rigid-body robot assumptions break down. Deep reinforcement learning trained in simulation, paired with tactile feedback that narrows positional uncertainty at the tool tip, enabled successful policy transfer without additional real-world tuning. The 67% recall on board-contact detection indicates the system sometimes misses contact events, which shapes the practical operating envelope for automated deployment.\n\n### Significance\n\nZME Science notes the robot is not ready to replace a sushi chef - cycle time remains slow under research safety limits, and failures occurred with the thinnest slices. The paper states the work 'represents a milestone in robotic manipulation of deformable, volumetric objects that may inspire and enable a wide range of other real-world applications.' The peer-reviewed npj Robotics publication marks an upgrade from the November 2025 arXiv preprint. Handling soft, variable-shape materials - food, fabric, biological tissue - remains a gap between lab robots and industrial deployments.\n\n## Scoring Rationale\n\nPublished in peer-reviewed npj Robotics (Nature), Sashimi-Bot demonstrates sim-to-real transfer and tactile-guided cutting of deformable objects - a genuine gap in robotic manipulation with applications in food processing and soft robotics. Score reflects solid domain-specific research with quantified test-set results; reduced from 6.6 to account for narrow application scope and 7-month lag from arXiv preprint to current press coverage.\n\nPractice interview problems based on real data\n\n1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.\n\n[Try 250 free problems](/problems)", "url": "https://wpnews.pro/news/sashimi-bot-uses-tactile-sensing-to-cut-salmon", "canonical_source": "https://letsdatascience.com/news/sashimi-bot-uses-tactile-sensing-to-cut-salmon-84047284", "published_at": "2026-06-20 00:38:16.568000+00:00", "updated_at": "2026-06-20 00:38:18.901886+00:00", "lang": "en", "topics": ["robotics", "machine-learning", "computer-vision", "ai-research"], "entities": ["Sashimi-Bot", "npj Robotics", "GelSight", "ZME Science", "Interesting Engineering", "Herland"], "alternates": {"html": "https://wpnews.pro/news/sashimi-bot-uses-tactile-sensing-to-cut-salmon", "markdown": "https://wpnews.pro/news/sashimi-bot-uses-tactile-sensing-to-cut-salmon.md", "text": "https://wpnews.pro/news/sashimi-bot-uses-tactile-sensing-to-cut-salmon.txt", "jsonld": "https://wpnews.pro/news/sashimi-bot-uses-tactile-sensing-to-cut-salmon.jsonld"}}