{"slug": "vention-s-jimmy-li-tells-lds-what-it-took-to-turn-a-robot-demo-into-a-factory", "title": "Vention's Jimmy Li tells LDS what it took to turn a robot demo into a factory system", "summary": "Vention's bin-picking robot required four months of core research on perception and motion planning to produce a trade-show demo, but deploying a complex cell at a customer factory took over a year of additional engineering, according to Dr. Jimmy Li, Vention's Director of Physical AI Research. Li told Lets Data Science that the team tuned its motion planner, built accurate collision models, changed the robot's finger material, added slight chamfers to receptacles, and overlapped planning with movement to hit the target cycle time. Li said Vention measures reliability through cycle time, percentage of parts dropped, and human interventions per hour, noting \"A missed attempt simply shows up as a longer overall cycle.", "body_md": "# Vention's Jimmy Li tells LDS what it took to turn a robot demo into a factory system\n\nIn written answers to LDS, Jimmy Li explains the perception, motion-planning and hardware work behind Vention's move from a bin-picking demo to a customer factory cell.\n\nA robot can pick a part successfully in a demonstration and still need substantial engineering before it is useful on a factory floor. In Vention's case, that work included changing the robot's finger material, refining collision models and selecting grasps that also allowed the part to be placed correctly.\n\nIn written answers to Lets Data Science, **Dr. Jimmy Li, Vention's Director of Physical AI Research**, said four months of core research on perception and motion planning produced a working trade-show demo for bin picking. Deploying a complex cell at a customer's factory then took **over a year of additional engineering**.\n\nThe task was to pick workpieces scattered in a bin and insert them neatly into receptacles. Seeing a part and finding a path to it were only the beginning. The system also had to meet the required cycle time and repeat the operation reliably.\n\nLi's answers to LDS add a concrete deployment example to the research-to-factory discussion in [Vention's September 9 interview about its Physical AI lab](https://vention.com/blogs/physical-ai/dr-jimmy-li-on-ventions-physical-ai-lab-1128). The timeline describes this development experience; it is not a stated average for current Vention installations.\n\n### A good pick also has to be a good placement\n\nLi said the team had to tune its motion planner and build accurate collision models for the surrounding scene, the robot and the workpiece. Those models describe the shapes the planner must account for when calculating movement.\n\nThe details become especially important at the bottom of a deep bin. A route must give the robot enough room to reach the part and retrieve it. Li said mature software utilities were needed to build and tune those representations.\n\nGrasp selection added another constraint. Holding an object securely is useful only if the robot can complete the next step with that grip.\n\nLi told LDS: \"Grasping also needs to be carefully tuned to not only cope with complex workpiece geometry, but also make downstream placement feasible.\"\n\nFor readers working with AI systems, that is a useful example of why success at one stage does not establish success for the whole task. A perception system can locate the right object, and a gripper can lift it, while the overall operation still struggles at insertion.\n\nSome of the consequential changes were physical. Li said changing the robot's finger material and adding slight chamfers, or beveled edges, to a receptacle had a large effect on consistency. His account connects software decisions with the surfaces the robot actually touches.\n\nThe team also overlapped planning with movement. Rather than waiting for a placement to finish before calculating the next pick, it performed that computation while the robot was placing the previous workpiece. Li said this helped the system reach its target cycle time, although he did not supply the target or the measured improvement.\n\n### Count the retries and the operator's time\n\nLDS asked how Vention measures reliability, including failed attempts and human intervention. Li described three operational metrics:\n\n- •**Cycle time:** the time between successful placements. The robot automatically retries a missed pick, so the extra attempt appears as a longer cycle. Li said his team does not track pick rate separately in this measurement approach.\n- •**Percentage of parts dropped:** how often a part falls onto the floor or somewhere the robot cannot recover it. Li identified poor grasps and misaligned placement as typical causes.\n- •**Human interventions per hour:** how often an operator must clear a fault during standard operation.\n\nLi said: \"A missed attempt simply shows up as a longer overall cycle.\"\n\nThis makes the definition of a successful cycle important. A recording of one clean pick cannot reveal how often the robot needed a second attempt or how much operator attention sustained the process. Li's chosen metrics account for retry time and the frequency of operator intervention, though an intervention count alone does not show how long each interruption lasts.\n\nThe answers explain how the team measures reliability, but do not report the results. The answers did not include cycle-time measurements, drop percentages, intervention counts, test duration or the number of placements observed. They also did not fully explain how resets and downtime enter the calculation, or provide results across shifts and changing conditions.\n\nReaders can therefore use the interview to understand what Vention measures. It does not establish a particular level of unattended reliability or a performance advantage over another system.\n\n### What zero-shot leaves for the engineer\n\nLi described a modular pipeline that combines foundation models for depth estimation, object segmentation and pose estimation. In plain terms, those stages help the system determine how far away objects are, separate them in an image and estimate their position and orientation.\n\nHe said the models can be prompted with text, images and computer-aided design models, allowing pick-and-place systems to be built without task-specific training.\n\nLi said: \"Our research shows that a robotic pick-and-place system can be built entirely zero-shot using a modular pipeline of foundation models, without any task-specific training.\"\n\n**Here, zero-shot refers to the absence of task-specific training.** It does not remove the planner tuning, collision geometry, grasp configuration and physical adjustments Li described. The distinction explains how a system can use pretrained models and still require substantial deployment work.\n\nVention's [public GRIIP SDK documentation](https://docs.vention.com/docs/griip-sdk-1) illustrates that configuration layer. It describes a Python library running on MachineMotion AI, with settings for the robot, tool, camera, collision geometry and parts. Its examples include part meshes and grasp files alongside application code.\n\nThe SDK documentation is publicly readable, but LDS has not confirmed general access to the SDK or tested it. Li's answers also did not identify a reproducible benchmark, model versions or an experiment an outside team could rerun to validate the zero-shot claim.\n\n### The next research step is more dexterous work\n\nLi said learning more intricate manipulation tasks through human teleoperation is on the horizon. In teleoperation, a person controls the robot to demonstrate a task. His team wants to reduce how much of that demonstration work is needed.\n\nHis proposed approaches include concentrating learning on the most complex parts of an operation, using reinforcement learning so the robot can learn through exploration and feedback, and training specialized manufacturing foundation models. These remain research directions in his answers.\n\nThat distinction is consistent with Vention's [public September 9 interview with Li](https://vention.com/blogs/physical-ai/dr-jimmy-li-on-ventions-physical-ai-lab-1128), which separates its current modular deployments from its plans for learned robot behavior.\n\n### What a deployment team can take from this\n\nLi's account suggests three practical questions for an AI or robotics team evaluating a demonstration:\n\n- •**Does the successful action enable the next one?** For this task, picking and placing have to work together. Evaluate the completed operation, including retries.\n- •**What had to change outside the model?** Ask about geometry, calibration, planning parameters, grippers and fixtures, as well as training data.\n- •**How much human attention does sustained operation require?** Ask for measured intervention rates, operating conditions and a clear account of downtime.\n\nThese questions apply the interview's lessons to deployment decisions. In Li's account, progress came from the interaction between perception, planning and the physical cell, including adjustments as tangible as finger material and receptacle edges. That is the work a successful demonstration can leave out.\n\n### Reporting note\n\nThis LDS Exclusive is based on Dr. Jimmy Li's written answers supplied directly to Lets Data Science through Vention. Vention's own public interview and SDK documentation provide background context. Deployment observations and the zero-shot claim are attributed to Li. LDS has not independently tested the robotic system; the answers did not include numerical reliability results or a reproducible benchmark.\n\n## Key Points\n\n- 1Li says four months of core perception and motion-planning research produced a demo; a complex customer cell required over a year of additional engineering.\n- 2Vention measures time between successful placements, unrecoverable drops and operator interventions, but supplied no numerical reliability results.\n- 3In Li's account, zero-shot means no task-specific training; collision models, grasp configuration, planner tuning and physical adjustments still matter.\n\n## Scoring Rationale\n\nOriginal written answers from Jimmy Li explain a bin-picking deployment, the engineering needed after a working demo, operational reliability metrics and the limits of zero-shot automation claims.\n\n## Sources\n\nOriginal reporting, with the public references used alongside it.\n\nLDS Exclusive\n\nReporting based on written answers given directly to Let's Data Science by **Dr. Jimmy Li, Director of Physical AI Research, Vention**.\n\n## View 1 more source\n\nPractice interview problems based on real data\n\n1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.\n\n[Try 250 free problems](https://letsdatascience.com/problems)", "url": "https://wpnews.pro/news/vention-s-jimmy-li-tells-lds-what-it-took-to-turn-a-robot-demo-into-a-factory", "canonical_source": "https://letsdatascience.com/news/vention-jimmy-li-tells-lds-how-a-robot-demo-reached-the-factory-floor-0239f886", "published_at": "2026-09-21 14:04:08+00:00", "updated_at": "2026-09-21 14:23:18.895740+00:00", "lang": "en", "topics": ["robotics", "artificial-intelligence", "ai-research"], "entities": ["Vention", "Jimmy Li", "Lets Data Science", "Physical AI lab"], "alternates": {"html": "https://wpnews.pro/news/vention-s-jimmy-li-tells-lds-what-it-took-to-turn-a-robot-demo-into-a-factory", "markdown": "https://wpnews.pro/news/vention-s-jimmy-li-tells-lds-what-it-took-to-turn-a-robot-demo-into-a-factory.md", "text": "https://wpnews.pro/news/vention-s-jimmy-li-tells-lds-what-it-took-to-turn-a-robot-demo-into-a-factory.txt", "jsonld": "https://wpnews.pro/news/vention-s-jimmy-li-tells-lds-what-it-took-to-turn-a-robot-demo-into-a-factory.jsonld"}}