Why Higher Success Rates Matter in Robotics

90% success rate in unseen environments. No new data, no fine-tuning. Autonomously. Most robots need retraining to work in new places. What if they didn’t? Robot Utility Models (RUMs) learn once and work anywhere... zero-shot. A team from NYU and Hello Robot built a set of general-purpose robot policies that can open drawers, pick up bags, and more, without needing extra training for each new home, setup, or lighting condition. Why RUMs matter ✅ Trained once and deployed in 25+ new environments without fine-tuning ✅ Uses a cheap handheld iPhone-based tool for fast, high-quality data collection ✅ Boosts performance with mLLM-based self-check and retry system ✅ Achieves 90 percent average success rate in real-world zero-shot tasks RUMs show that smart data collection and multi-modal learning can unlock truly general robotics. Try it yourself: Code, models, and data are all open source at https://lnkd.in/dNH_HCat.

The $437B robotics revolution is failing at execution, not tech. Hardware works. Software works. Deployment fails. Companies burn millions on robots they never use. Here's why: A CEO invests in cutting-edge robotics. Perfect pilot test. Scale-up? Total collapse. This pattern threatens the entire industry's future: • $2M robot arms gathering dust in manufacturing plants • Food prep automation sitting unused in restaurant storage • Warehouses reverting to manual processes after failed implementation The tech works flawlessly, with 99.8% defect detection accuracy and 24/7 operation. But something breaks between lab testing and real-world execution. For every $1 million in robotics hardware, companies waste $3-5 million on failed implementations. They chase better tech when their problem is execution support. A food manufacturer spent $7M on automation that worked in testing but failed at scale. No one was there to execute the transition. An e-commerce giant saw picking accuracy drop from 99% in pilots to 87% in deployment. The critical success factor? On-site implementation experts during rollout, delivering 4x higher success rates than remote support. This insight drove the creation of RobotLAB: Local robotics experts who master your implementation instead of shipping hardware and hoping for the best. The results: 65% faster deployment and 93% higher first-year ROI. Our teams handle: • Staff training • Set up execution • Systems integration Manufacturing clients achieve full deployment in 3 weeks instead of 4-6 months. Logistics operations reach 98% accuracy within days, not months. The robotics revolution demands elite execution in the critical "last mile" of implementation. If you're considering robotics or struggling with implementation, I can help. I wrote "Our Robotics Future" as a practical guide for business leaders. DM me for a copy or to discuss how our teams can upgrade your robotics execution.

What if robots could react, not just plan? A good read: https://lnkd.in/gEGSp_5U This paper proposes a Deep Reactive policy (DRP), a visuo-motor neural motion policy designed for generating reactive motions in diverse dynamic environments, operating directly on point cloud sensory input. Why does it matter? Most motion planners in robotics are either: Global optimizers: great at finding the perfect path, but they are way too slow and brittle in dynamic settings. Reactive controllers: quick on their feet, but they often get tunnel vision and crash in cluttered spaces. DRP claims to bridge the gap. And what makes it different? 1. IMPACT (transformer core): pretrained on 10 million generated expert trajectories across diverse simulation scenarios. 2. Student–teacher fine-tuning: fixes collision errors by distilling knowledge from a privileged controller (Geometric Fabrics) into a vision-based policy. 3. DCP-RMP (reactive layer): basically a reflex system that adjusts goals on the fly when obstacles move unexpectedly. Results are interesting for real-world evaluation: Static environments: Success Rate: DRP 90% | NeuralMP 30% | cuRobo-Voxels 60% Goal Blocking: Success Rate: DRP 100% | NeuralMP 6.67% | cuRobo-Voxels 3.33% Goal Blocking: Success Rate: DRP 92.86% | NeuralMP 0% | cuRobo-Voxels 0% Dynamic Goal Blocking: Success Rate: DRP 93.33% | NeuralMP 0% | cuRobo-Voxels 0% Floating Dynamic Obstacle: Success Rate: DRP 70% | NeuralMP 0% | cuRobo-Voxels 0% What stands out from the results is how well DRP handles dynamic uncertainty, the very scenarios where most planners collapse. NeuralMP, which relies on test-time optimization, simply can’t keep up with real-time changes, dropping to 0 in tasks like goal blocking and dynamic obstacles. Even cuRobo, despite being state-of-the-art in static planning, struggles once goals shift or obstacles move. DRP’s strength seems to come from its hybrid design: the transformer policy (IMPACT) gives it global context learned from millions of trajectories, while the reactive DCP-RMP layer gives it the kind of “reflexes” you normally don’t see in learned systems. The fact that it maintains 90% success even in cluttered or obstructed real-world environments suggests it isn’t just memorizing scenarios; it has genuinely learned a transferable strategy. That being said, the dependence on high-quality point clouds is a bottleneck. In noisy or occluded sensing conditions, performance may degrade. Also, results are currently limited to a single robot platform (Franka Panda). So this paper is less about replacing classical planning and more about rethinking the balance between experience and reflex. 

Medtronic conducted 137 real-world surgical tests with its Hugo robot—performing prostate, kidney, and bladder repairs—with results surpassing physicians' expectations. Complication rates were remarkably low: just 3.7% for prostate surgeries, 1.9% for kidney procedures, and 17.9% for bladder operations, all exceeding the safety benchmarks established through years of research. The robot achieved a 98.5% success rate, far surpassing the 85% target—meaning it didn’t just pass the tests but effectively set a new standard. Out of 137 procedures, only two required conversion to traditional surgery (one due to a robotic malfunction and the other due to the patient's complex condition). This doesn’t mean robots will replace surgeons tomorrow, but it does suggest your next doctor might have a very expensive metal partner by their side. メドトニック社はそのHugoロボットに対して137回の実際の手術テスト（前立腺、腎臓、膀胱の修復）を実施し、結果は医師の予想を上回るものでした。 合併症発生率は極めて低く：前立腺手術でわずか3.7%、腎臓手術で1.9%、膀胱手術でも17.9%と、いずれも長年の研究で設定された安全目標を達成しました。 このロボットの成功率は98.5%に達し、85%という目標を大きく上回りました——これは単にテストを通過しただけでなく、事実上新たな基準を設定したことを意味します。 137例の手術中、従来の手術に切り替えが必要だったのはわずか2例（1例はロボットの故障、もう1例は患者の状態が複雑だったため）でした。 これは「ロボットが明日から外科医に取って代わる」という意味ではありませんが、次に診てくれる医師の横に非常に高価な金属の相棒がいる可能性は大いにあると言えるでしょう。

Physical Intelligence (Pi) just dropped Pi-0.5 — a new foundation model for robots that tackles general tasks like cleaning & laundry. Trained via VLM + action tokenization, it performs well even in unseen environments. Trained on 400+ hours of diverse mobile manipulation data and multi-modal sources (web, vision, language, and robot demos), π0.5 can perform complex real-world tasks like cleaning tables, folding laundry, or loading dishwashers—even in unseen homes with new layouts and unfamiliar objects. How it works: π0.5 is co-trained on a rich mix of heterogeneous data—robot demos, image-text tasks (like captioning or object detection), and natural language instructions. It uses a dual-pathway architecture: 🔠 Discrete token inference for high-level planning (“put dishes in sink”) 🔁 Continuous decoding via flow-matching for low-level motor control (e.g., joint movements) The model outputs a sequence of 50-step “action chunks” (~1 second each), chaining high-level reasoning to low-level physical action. It can even transfer behaviors from simpler robots or simulations to more complex hardware. 📊 Performance: In evaluations across new environments, π0.5 achieved strong generalization: - Up to 90%+ task success rates on multi-step cleaning tasks - Significant gains from including diverse training sources (e.g., adding web data improved novel object recognition in out-of-distribution tests). - Generalization improved steadily with training on just 100 distinct environments, nearing the performance of models trained directly on test scenes. This matters because current robots are often limited to fixed settings—like warehouses. π0.5 shows real promise in dynamic, cluttered, everyday spaces like homes, offices, and hospitals. In short, π0.5 doesn’t just learn tasks—it learns how to learn tasks in new places. That’s a massive leap toward general-purpose home and service robotics.