> ## Content Index > Fetch the complete content index at: https://www.japanlayman.com/llms.txt > Use this file to discover other available public pages before exploring further. # Can you teach a robot not to squash a shrimp? - URL: https://www.japanlayman.com/can-you-teach-a-robot-not-to-squash-a-shrimp/ - Published: 2026-09-06T12:18:21.000Z - Updated: 2026-09-06T12:18:21.000Z - Author: Japan Layman *A Hokkaido University startup is trying to teach robots not just where to move, but how hard to push, hold and squeeze.* Industrial robots are very good at doing the same thing over and over again. Give one a precisely positioned metal component and a carefully programmed movement, and it can repeat the job thousands of times. Give it something soft, slippery and slightly different every time, and life becomes more difficult. This is one reason seafood processing remains surprisingly dependent on human hands. Fish, shellfish and roe vary in shape, size and firmness. Some jobs require enough force to pick something up, but not so much that you damage it. A startup from Hokkaido University called [Real Touch](https://real-touch.studio.site/?ref=japanlayman.com) is trying to teach robots that part. One of its demonstrations involves **tarako**, soft sacs of pollock roe. The robot grips and moves them without simply crushing them between its fingers. More recently, the company has been paired with a land-based shrimp producer in Shizuoka to work on something more practical: automating the sorting, weighing and packing of vannamei shrimp. The interesting bit is not really the shrimp. It is the idea that physical AI — AI that has to perceive and act in the physical world — may need to learn something humans barely think about. **How hard am I touching this?** **Robots are good at knowing where they are** Traditional industrial automation works beautifully when the world around the robot is controlled. Parts arrive in predictable positions. Their dimensions are known. The robot has been programmed for a particular movement. If the same situation appears again, the same motion can be repeated with extraordinary precision. Food is less cooperative. [Japan's Fisheries Agency](https://www.jfa.maff.go.jp/j/kikaku/wpaper/r01%5Fh/trend/1/t1%5Ff2%5F6.html?utm%5Fsource=chatgpt.com) has pointed out that fish-processing automation has historically lagged other manufacturing sectors partly because the things being handled are inconsistent in shape and size. That creates two related problems. The robot needs to work out **where** the object is. Then, once it touches it, it needs to know **how to interact with it**. A human hand handles this almost unconsciously. If something begins to slip, your grip tightens. If a strawberry, piece of fish or other soft object starts deforming, you ease off. You constantly adjust force according to what you feel. Real Touch's approach is to capture some of that physical interaction as data. Its [NEDO project](https://nep.nedo.go.jp/selected/aaaa6833-cab1-4f6b-9655-80a4bfd9d1aa?utm%5Fsource=chatgpt.com) combines two technological ingredients: **Real Haptics**, a force-haptics technology with roots at Keio University, and AI trained using digitised force information alongside movement data. The objective is to let somebody physically demonstrate a task and have the robot learn more than the trajectory through space. It also learns something about the forces involved. ![](https://storage.ghost.io/c/32/16/3216a63f-8c30-4d6c-a71f-2837fc166f59/content/images/2026/09/Pixel-Fox-Haptic-Robot-Interface.png) **What does the robot actually feel?** Real Haptics is easier to understand if you imagine a person remotely controlling a robot. Move the controller in your hand and the remote robot follows the movement. But information also travels in the other direction. If the robot presses against something, the force pushing back on the robot is transmitted back to the controller. The operator can feel resistance from an object they are not physically touching. Real Haptics does this by synchronising both **position and force** between the two sides. [Motion Lib](https://www.motionlib.com/product/abc-core/?utm%5Fsource=chatgpt.com), the Keio University spinout commercialising the technology, describes this as bidirectional transmission of force-haptic information. That creates another useful possibility. The system can record not just where the operator moved, but how much force they used while doing it. Human actions that normally exist only as experience in somebody's hands can become data. Those data can then be used to reproduce the movement automatically. Real Touch is taking that idea and adding learning on top. Founder **Shun Maki** was already researching haptic-feedback robotics at Hokkaido University before the company existed. In 2024, he and colleagues presented [a teleoperation robot incorporating haptic feedback](https://www.jstage.jst.go.jp/article/jsmermd/2024/0/2024%5F2P2-D08/%5Farticle/-char/en?utm%5Fsource=chatgpt.com) at the Japan Society of Mechanical Engineers' Robotics and Mechatronics conference. That same year, his project was selected for Hokkaido's [university GAP Fund programme](https://hsfc.jp/gap-fund-2024/?utm%5Fsource=chatgpt.com) under a rather good title: **“A general-purpose robot that anyone can teach, equipped with force sensation, and reform of the seafood industry.”** Real Touch was incorporated in August 2025\. It has since also been selected for NEDO support to develop the technology further. ![](https://storage.ghost.io/c/32/16/3216a63f-8c30-4d6c-a71f-2837fc166f59/content/images/2026/09/Fox-Demonstrates--Robot-Learns.png) **The five-minute robot** This is where the proposition becomes more ambitious. Real Touch says an operator can teach its robot a new manual task in around **five minutes**. Rather than a robotics engineer programming each action, somebody demonstrates the job while the system records movement and force data. The company also says the robot can use that learning when the objects it encounters vary. Its tarako demonstration is the memorable example. In a [January 2026 interview with ASCII Startup](https://ascii.jp/elem/000/004/353/4353984/?utm%5Fsource=chatgpt.com), Maki said that, to his knowledge, no other robot could pick up and move pieces of tarako across different shapes without crushing them in the same way. That is an interesting claim. It is not something that has been independently established. There also isn't yet enough public performance data to know what “five minutes” would mean on a real production line. The public sources reviewed for this article do not provide figures for cycle time, success rates, failure rates, the range of variation it can tolerate or how often somebody has to intervene. [NEDO's description of the current project](https://www.nedo.go.jp/content/800037310.pdf?utm%5Fsource=chatgpt.com) gives a useful reality check. It says Real Touch is developing a seafood-processing **MVP with the aim of progressing to paid proof-of-concept projects**. More detailed NEDO material says the lab-tested MVP still needs improvements in speed, accuracy, robustness and usability before progressively moving through factory PoCs. In other words, this is still early. **Why seafood is such a good test** Hokkaido is an obvious place to try this. Food manufacturing matters to the regional economy, while [labour availability is a continuing problem](https://www.jfa.maff.go.jp/j/kikaku/wpaper/r06%5Fh/trend/1/t1%5F2%5F8.html?ref=japanlayman.com). [Hokkaido's regional METI bureau](https://www.hkd.meti.go.jp/hokcf/20250527/index.htm?utm%5Fsource=chatgpt.com) has specifically been pushing labour-saving investment and factory automation in food manufacturing, including robots and automated machinery. But seafood is also a particularly unforgiving automation problem. Products can be wet, fragile, slippery and irregular. Their shape changes from one item to the next. A plant may also handle different products or shorter production runs rather than making millions of identical components. Conventional automation becomes harder to justify if engineers have to spend substantial time setting up the robot every time the job changes. That is the bigger idea underneath Real Touch's five-minute claim. Can you make sophisticated automation worthwhile in factories where **variation itself has historically made automation difficult?** If so, seafood might only be the beginning. **From tarako to an NTT shrimp operation** The recent breadcrumb that brought Real Touch back onto the radar came from **Iwata in Shizuoka**. In July 2026, [Iwata selected Real Touch for two projects](https://www.city.iwata.shizuoka.jp/sangyou%5Fbusiness/kigyou%5Fshien/1016735.html?utm%5Fsource=chatgpt.com) under a programme matching startups with local companies. One is with Matsushita Kogyo, an automotive casting-mould specialist that has more recently moved into robot systems integration. The companies are working on a prototype transport cart intended to help put Real Touch's robot into production environments. The other involves shrimp. **Kaiko Yukinoya** produces vannamei shrimp using a fully closed recirculating land-based aquaculture system in Iwata. Real Touch is working with it on a system intended to automate parts of the shipping process: **sorting, weighing and packing**. The programme describes this specifically as [robot-system development and demonstration work](https://prtimes.jp/main/html/rd/p/000000194.000099287.html?ref=japanlayman.com). But Kaiko is more interesting than the phrase “local shrimp farm” suggests. It began as a Kansai Electric-led venture before being acquired in 2024 by **NTT Green & Food**, a joint venture between NTT and Regional Fish, a Kyoto University-origin fisheries startup. [NTT Green & Food acquired all of Kaiko's equity in August 2024](https://www.ntt-green-and-food.com/ir-news/499/?utm%5Fsource=chatgpt.com). Kaiko is now a wholly owned subsidiary of NTT Green & Food. Its facility has a listed production capacity of around **80 tonnes of shrimp a year**. NTT Green & Food also operates a separate Iwata shrimp plant with listed capacity of around **110 tonnes a year**. Both capacities are listed on [NTT Green & Food's current facility directory](https://www.ntt-green-and-food.com/partner-region-list/?utm%5Fsource=chatgpt.com). An [NTT technical publication](https://journal.ntt.co.jp/article/34187?utm%5Fsource=chatgpt.com) describes the two operations together as Japan's largest land-based shrimp producer, with roughly 200 tonnes of annual production capacity. So Real Touch is not simply entering a demonstration with a tiny experimental aquaculture farm. A very early Hokkaido startup has entered a development-and-demonstration project with a shrimp producer inside an NTT-backed aquaculture business operating at meaningful industrial scale. That still needs careful wording. The Iwata project is a **development and demonstration project**. There is no evidence that NTT Green & Food has adopted Real Touch commercially, or that the robot is already operating routinely on a production line. But it makes the experiment worth paying attention to. The tarako demonstration shows what the technology might do. The shrimp project starts testing whether there is a useful industrial job for it. ![](https://storage.ghost.io/c/32/16/3216a63f-8c30-4d6c-a71f-2837fc166f59/content/images/2026/09/Pixel-Art-Shrimp-Processing-Facility.png) **How unusual is this?** Real Touch is not alone in trying to make robots better at physical contact. Around the world, robotics researchers are combining tactile sensing, force information, human demonstrations and imitation learning to improve tasks that are difficult to solve using vision alone. Recent work presented at [Robotics: Science and Systems](https://www.roboticsproceedings.org/rss21/p052.html?utm%5Fsource=chatgpt.com), for example, combines visual and tactile feedback with imitation learning for contact-rich manipulation. Soft robotics researchers have also been working on compliant grippers for fruit, food and other fragile objects for years. Handling these objects quickly without damaging them remains an active research problem rather than a solved one. More recently, improvements in AI have made it possible to learn increasingly complicated physical behaviours from demonstrations and large datasets. American robotics company **Dexterity**, for example, is developing its own approach to Physical AI for logistics. The company says its [Instinct system uses force-guided manipulation in production](https://dexterity.ai/blog/instinct?utm%5Fsource=chatgpt.com), and in July 2026 [Dexterity and FedEx announced an expanded deployment](https://dexterity.ai/blog/fedex-hagerstown-physical-ai-deployment?utm%5Fsource=chatgpt.com) of its autonomous trailer-loading systems. So Japan has not uniquely discovered that robots need touch. What makes Real Touch interesting is narrower. It is trying to combine that broader movement in robotics with a problem that matters in Japanese seafood processing: **how do you automate skilled physical work when the products are variable, production runs may be relatively small, and conventional automation can be too expensive or inflexible?** Seafood processing is an unusually good place to find out. **What would change the story?** The next phase should tell us much more than another impressive demonstration video. The first thing to watch is whether one of Real Touch's development projects becomes a **paid production PoC**, and then whether that becomes an actual deployment. Then come the boring numbers that matter enormously. How quickly can it perform a task? How often does it fail? How much variation can it handle without retraining? How long does changing a task really take? Can factory workers actually teach it without a robotics engineer? How does the complete system compare in price with conventional automation or simply continuing to use human labour? Food processing creates another set of practical questions. Equipment has to survive cleaning, moisture and repetitive industrial use. Contact surfaces also have to fit within food-hygiene requirements. An interesting manipulation algorithm is only one piece of an actual food-production machine. Public material does not yet answer most of these questions. That is precisely why the Iwata shrimp project is worth following. Real Touch does not need to build a robot that can do everything. If it can make one awkward, variable production job reliably economical to automate, that would already be interesting. **The rabbit hole: the same idea is loading explosives into tunnels** ![](https://storage.ghost.io/c/32/16/3216a63f-8c30-4d6c-a71f-2837fc166f59/content/images/2026/09/Pixel-Fox-Tunneling-Control-Room.png) The technology underneath Real Touch has taken a much stranger route elsewhere. Keio University and Japanese construction company **Obayashi** have been applying Real Haptics to one of the more dangerous manual jobs in tunnel construction: putting explosives into holes drilled in the rock face before blasting. The problem again involves force. Explosives, detonators and their connecting lines require delicate handling. The tunnel face is also precisely where you would prefer workers not to stand because freshly excavated rock can fall. In January 2026, [Obayashi and Keio reported a field trial](https://www.obayashi.co.jp/news/detail/news20260120%5F1.html?utm%5Fsource=chatgpt.com) at the **New Sasago Tunnel** project in Yamanashi. According to Obayashi, loading work previously performed by five people directly beneath the tunnel face was carried out by **one operator around 50 metres away**. Real Haptics allows that operator to control the machinery while feeling the forces generated as explosives are inserted into the drilled holes. Movement and force data gathered during remote operation can also be used as the basis for automation. The project has another connection to the Real Touch story. It has been developed under [NEDO's Young Researcher Support Program](https://www.nedo.go.jp/news/press/AA5%5F101795.html?ref=japanlayman.com). So two completely different applications of force-haptics technology are being pulled towards the real world through NEDO programmes. One is trying to automate delicate seafood processing. The other is trying to remove workers from directly beneath a tunnel face while explosives are being loaded. They are not the same robot and Real Touch is not involved in the Obayashi project. But the common technological idea is striking. If a machine can record, transmit and reproduce the forces involved when a skilled person interacts with something, physical know-how that normally lives in somebody's hands begins to look more like data. One branch is trying not to damage fish roe. Another is trying to keep people away from explosives and falling rock. The common problem is something our own hands solve so naturally that it is easy to forget how difficult it is to teach a machine: **not just what movement to make, but how hard to make it.** **Sources & Further Reading** [**NEDO — Real Touch project**](https://www.nedo.go.jp/content/800037310.pdf?utm%5Fsource=chatgpt.com) Probably the best source for separating the company's ambition from its current development stage. NEDO describes the combination of Real Haptics and force-data learning, the lab-tested MVP, the work still needed on speed, accuracy and robustness, and the goal of reaching paid PoCs. [**Hokkaido Future Creation Startup Ecosystem / HSFC**](https://hsfc.jp/gap-fund-2024/?utm%5Fsource=chatgpt.com) The university commercialisation programme that supported Shun Maki's project before Real Touch was incorporated, including the original GAP Fund project focused on force-enabled robots and seafood processing. [**J-STAGE / Japan Society of Mechanical Engineers**](https://www.jstage.jst.go.jp/article/jsmermd/2024/0/2024%5F2P2-D08/%5Farticle/-char/en?utm%5Fsource=chatgpt.com) Maki and colleagues' 2024 work on a teleoperated robot with haptic feedback, useful for seeing the research underneath the later startup. [**ASCII Startup**](https://ascii.jp/elem/000/004/353/4353984/?utm%5Fsource=chatgpt.com) A detailed interview and demonstration profile of Real Touch, including the tarako handling, five-minute teaching proposition and Maki's novelty claim. These are useful for understanding what the company says the technology can do, rather than as independent validation of its performance. [**Motion Lib / Keio University-derived Real Haptics technology**](https://www.motionlib.com/company/?utm%5Fsource=chatgpt.com) The clearest route into how Real Haptics synchronises position and force, transmits force information bidirectionally and turns physical interaction into data that can be recorded and reproduced. [**Iwata City**](https://www.city.iwata.shizuoka.jp/sangyou%5Fbusiness/kigyou%5Fshien/1016735.html?utm%5Fsource=chatgpt.com) The official record of the current projects with Matsushita Kogyo and Kaiko Yukinoya. [**Iwata programme project descriptions**](https://prtimes.jp/main/html/rd/p/000000194.000099287.html?ref=japanlayman.com) Useful additional detail on what the two Real Touch collaborations are actually trying to build, including shrimp sorting, weighing and packing. [**NTT Green & Food — Kaiko acquisition**](https://www.ntt-green-and-food.com/ir-news/499/?utm%5Fsource=chatgpt.com) and [**current Iwata facilities**](https://www.ntt-green-and-food.com/partner-region-list/?utm%5Fsource=chatgpt.com) Background on Kaiko Yukinoya, its acquisition by NTT Green & Food and the scale of the wider Iwata shrimp operation. [**Japan Fisheries Agency**](https://www.jfa.maff.go.jp/j/kikaku/wpaper/r01%5Fh/trend/1/t1%5Ff2%5F6.html?utm%5Fsource=chatgpt.com) and [**Hokkaido Bureau of METI**](https://www.hkd.meti.go.jp/hokcf/20250527/index.htm?utm%5Fsource=chatgpt.com) Context on why seafood and food processing are difficult to automate, including inconsistent products, employee shortages and current efforts to introduce robots and automated machinery. [**Obayashi — New Sasago Tunnel trial**](https://www.obayashi.co.jp/news/detail/news20260120%5F1.html?utm%5Fsource=chatgpt.com) and [**NEDO — Real Haptics explosives-loading programme**](https://www.nedo.go.jp/news/press/AA5%5F101686.html?utm%5Fsource=chatgpt.com) The primary sources for the tunnel rabbit hole, including the 50-metre remote operating distance, reduction from five workers at the tunnel face to one operator, and the wider NEDO programme. [**Robotics: Science and Systems — visual/tactile imitation learning**](https://www.roboticsproceedings.org/rss21/p052.html?utm%5Fsource=chatgpt.com) Useful global context showing that tactile feedback and learning from demonstrations are active international robotics research areas rather than something unique to Real Touch or Japan.