Today on The Robot Beat: humanoids move from demo reels to enterprise pilots at Accenture/SAP/Vodafone and UBTECH/ROSSMANN as AgiBot crosses 10,000 units shipped, Bessemer calls robotics structurally under-invested as 90% of humanoids ship from China, and Qualcomm launches a $300 Jetson challenger while Tesla details Optimus Gen 3's tendon-driven hand.
Accenture, SAP, and Vodafone Procure & Connect completed a pilot deploying humanoid robots for autonomous visual inspection in a Duisburg, Germany warehouse. Humanoids were trained in digital twins via Accenture's Robot Brain stack and integrated directly with SAP's Extended Warehouse Management system for real-time reporting — arriving 72 hours after the Siemens/NVIDIA Erlangen eight-hour shift trial.
Why it matters
The integration pattern is the story: Robot Brain as the embodied AI layer, SAP EWM as the enterprise system of record, and a Big Four systems integrator as the deployment channel. This is how enterprise buyers actually procure — not direct from a humanoid OEM. The winning platforms may be those with clean APIs into SAP/Oracle/Siemens stacks rather than the best locomotion videos.
Bulls read this as Big Four consulting pipelines opening, which typically precede 18–24 month enterprise rollouts. Skeptics note 'pilot for visual inspection' must justify the humanoid form factor over a fixed camera plus VLM — mobility and manipulation need to do work, not just looking at things.
Extending yesterday's AgiBot Partner Conference coverage: AgiBot confirmed its 10,000th humanoid shipped, with 5,000 units in the last 90 days versus two years for the prior 5,000. New platforms unveiled include the A3 humanoid, G2 Air mobile manipulator, OmniHand 3 Ultra-T dexterous hand, D2 Max quadruped, eight foundation models, and MEgo — a data-collection system that trains without deployed hardware.
Why it matters
The Q1 production curve is the most aggressive humanoid manufacturing acceleration disclosed publicly, and sets AgiBot as the volume reference against which Tesla Optimus, Figure, and Apptronik are now measured. MEgo's training-without-hardware approach is the new detail — it signals a platform strategy (hardware + models + data pipeline) rather than OEM-only. This directly challenges the Bessemer under-investment thesis by showing a self-funding Chinese player operating outside Western VC norms.
Production volume still doesn't equal deployed hours — the Figure/Tesla deployed-hours-weighted comparison from April 19 remains the harder metric. Chinese humanoid valuations remain priced as hardware plays until AgiBot shows durable multi-shift operations at named end-users.
Skild AI acquired Zebra Technologies' robotics automation business, absorbing the Symmetry Fulfillment orchestration platform. The combined offering pairs Skild Brain — a generalist embodied AI intended to control any robot across any task — with Symmetry's multi-vendor fleet orchestration. Skild's pitch is end-to-end warehouse automation without requiring customers to restructure operations around a specific robot hardware spec.
Why it matters
This is the first real test of whether a hardware-agnostic AI stack plus existing fleet orchestration can outcompete vertically integrated humanoid OEMs in warehouse automation. Zebra's Symmetry customer base gives Skild immediate distribution into live deployments — a harder asset to replicate than the Brain itself. For entrepreneurs, the structural signal is that robotics M&A has moved from hardware tuck-ins to AI-stack-acquires-incumbent-integrator, consolidating the 'control layer' over heterogeneous robot fleets.
Bulls argue Skild is assembling the equivalent of an Android for warehouse robotics — one generalist policy, many hardware form factors. Skeptics point out that 'omnibodied' generalist policies have not yet demonstrated parity with task-specific models on measured throughput, and that Zebra's PSS divestiture to Brady (also announced this week) signals Zebra itself is de-prioritizing this line — raising questions about the quality of the assets Skild actually acquired.
UBTECH appointed Terra Robotics as exclusive DACH distributor with initial deployment of dozens of Walker S humanoids at ROSSMANN logistics centers — the first disclosed retail-logistics customer beyond BYD/Geely automotive lines. UBTECH disclosed 2025 humanoid revenue of 821M yuan, a 22x jump from 2024, on total revenue of 2.001B yuan (+53.3% YoY) and 1,079 cumulative humanoid units shipped.
Why it matters
A concrete Chinese-humanoid-in-Europe commercial deployment at a named retailer directly counters the 'Chinese humanoids can't sell outside China' thesis, and pairs with today's Bessemer 90%-shipment finding. UBTECH's distributor-channel approach trades gross margin for EU regulatory/compliance bypass — watch for renewal or expansion in six months as the real performance signal.
'Dozens of units' is a starter deployment. The distributor model is strategically sound given EU hardware-import friction, but limits UBTECH's direct customer relationship — a vulnerability if Terra Robotics pivots suppliers.
Tesla revealed Optimus Gen 3 details: a mechanically tendon-driven hand with four degrees of freedom per finger, heavy actuators relocated to the forearm to reduce distal mass, and ~50 actuators total. Musk framed hand perfection as a Cybertruck/Model-X-class engineering problem.
Why it matters
Tendon routing through the wrist aligns with this week's dexterous-hand price halving and Schaeffler's integrated actuator push, but Tesla still has no external customer deployments — against Figure 03's 1,250+ BMW hours. Architectural announcements have historically outpaced Optimus shipped performance; watch for Gen 3 appearing in Tesla's own Fremont/Giga Texas lines before customer pilots.
Proximal actuator relocation addresses known inertia and control-bandwidth problems, but cable wear and calibration drift have sunk multiple prior tendon-driven programs. The asymmetry with Figure's disclosed deployment hours remains the central credibility gap.
Zoomlion Heavy Industry unveiled Robot Ops at Hannover Messe 2026, a development platform for embodied intelligence that unifies DevOps, DataOps, and AgentOps across full lifecycle from data collection to deployment. Multi-robot collaboration scenarios were demonstrated with the Z1 humanoid alongside Zoomlion's Industry 5.0 manufacturing portfolio. The company claims ~50% iteration-efficiency improvement for robotics developers.
Why it matters
Construction-machinery OEMs entering the humanoid/embodied-AI platform market is a meaningful tell — Zoomlion has the same structural advantages cited for Schaeffler and Infineon: captive heavy-duty customers, industrial sales channels, and actuator/power-electronics expertise. Robot Ops is effectively a competitor to Accenture's Robot Brain and AgiBot's MEgo at the fleet-ops layer.
Ops-platform wars typically resolve toward one or two standards (Kubernetes, ROS). A Zoomlion-branded Robot Ops may struggle to achieve cross-OEM adoption versus a neutral player, but could become the de facto standard within the Chinese heavy-machinery ecosystem. Worth watching whether Sany, XCMG, or other peers adopt or compete.
Bosch and ECOVACS jointly launched the world's first built-in vacuum/mop robot integrating directly into kitchen cabinetry, with a hidden wet-dry service station connected to household water and power. The 84mm-tall robot delivers 20,000 Pa suction; charging, self-cleaning, and hot-water mop sanitization occur entirely behind cabinet doors.
Why it matters
This is a genuine form-factor shift — infrastructure-integrated appliance rather than mobile standalone device — directly addressing the #1 design objection from high-end homeowners. Contrasts sharply with yesterday's Narwal Flow 2 (60°C hot water, VLM vision, 31,000 Pa) and positions Bosch+ECOVACS against the Chinese pure-play OEM race on specs with a kitchen-channel distribution bet instead.
The competitive frame is clear: Chinese pure-plays race on suction/vision/hot-water (Narwal, Roborock, Dreame); Bosch+ECOVACS bets on integration and kitchen-channel distribution. Different go-to-market, potentially different buyer — the parallel to built-in dishwashers is instructive for whether this becomes a category or a one-off.
Dreame aired a $10M Super Bowl ad announcing aggressive diversification into hypercars, humanoids, satellites, and consumer electronics, with CEO Yu Hao positioning himself as 'China's Elon Musk.' A San Francisco launch event April 27 introduces premium TVs, refrigerators, washer-dryers, and an AI laundry robot to the US market.
Why it matters
The Chinese floorcare category is saturating (Narwal, Roborock, Ecovacs, Xiaomi all in premium tier), and Dreame is betting brand + capital + manufacturing relationships transfer across categories. The AI laundry robot is the most technically interesting piece — laundry manipulation remains unsolved even for Figure and 1X, and a credible consumer laundry robot would be category-defining. The April 27 SF event is the near-term datapoint.
Samsung precedent bulls: vertical integration across appliances, TV, and mobile worked via chaebol capital and semiconductor ownership. LeEco precedent bears: simultaneous expansion into cars + TVs + phones + content collapsed on cashflow and focus.
Google released Gemini Robotics-ER 1.6, a reasoning-centric update for industrial robotics adding spatial reasoning improvements, multi-view/multi-camera perception, instrument reading (gauges, meters), and explicit task-success detection. It is positioned as a high-level reasoning/planning layer above low-level control policies rather than end-to-end control.
Why it matters
The instrument-reading and task-success-detection additions address specific gaps in prior VLM-for-robotics releases required for unsupervised industrial deployment. This doubles down on the two-system (reasoner + actor) pattern, directly countering OneTwoVLA's unified architecture (covered April 21, 87% long-horizon success). The field is bifurcating: Google bets the reasoning layer is best upgraded independently of low-level control; OneTwoVLA bets unified models eliminate the latency and state-sync costs of separation.
Building on yesterday's VinDynamics-Schaeffler announcement: Schaeffler's second humanoid OEM deal this week adds Southeast Asia to its engagement footprint via the Hermes Award-winning actuator platform. Companion product news: Schaeffler launched EMA-50 and EMA-60 compact electromechanical linear actuators (1–2 ton range) via its EWELLIX portfolio.
Why it matters
Two humanoid OEM deals across Europe and Asia in one week confirms Schaeffler is actively consolidating at the actuator/gearbox supply layer — the same dynamic as Harmonic Drive's 500K→800K gearbox ramp and dexterous-hand price halving, both covered April 21. With component order books full through 2027, actuator sourcing is now a strategic priority equal to software talent for any humanoid startup.
Melexis and OYMotion integrated Tactaxis — Melexis's 3D magnetic tactile sensor — into industrialized fingertip modules for robotic hands, moving from prototype to production-ready form. True 3D force, position, and contact-interaction sensing is the intended capability, with a live demo planned at Hannover Messe.
Why it matters
Production-grade 3D tactile modules in a fingertip form factor are what enable algorithms like DexMove (covered April 19) to be deployed outside the lab, and directly addresses one of the three hardware bottlenecks TI's German Aguirre flagged this week alongside deterministic control and GaN power. The adoption gating factor is not the sensor but the data pipeline — VLA research labs need large tactile-labeled manipulation datasets, and Melexis/OYMotion don't provide that.
Texas Instruments' systems manager for robotics German Aguirre mapped the hardware bottlenecks between humanoid prototype and production: reliability-at-scale in BMS and power stages, deterministic real-time control via purpose-built MCUs, GaN power stages for compact dexterous hand control, and high-resolution multimodal sensing with sensor fusion.
Why it matters
Complements the Infineon CEO's $96B humanoid chip TAM call (April 19) and this week's Schaeffler actuator and Melexis tactile stories — automotive-grade semiconductor suppliers are converging on humanoid hardware as the next growth segment with specific technical bets. Consistent with Microchip's CLB-based PIC MCU launch this week, which bundles CPLD-class logic onto sub-$1 MCUs for exactly these motor-control and real-time safety scenarios. The chip vendors are arriving more prepared than the software stack vendors.
Bessemer Venture Partners published a report arguing robotics is dramatically under-capitalized: only 42 robotics companies raised >$30M over five years versus 745 software companies, even as 90% of 2025 humanoid shipments came from China. Bessemer forecasts 50x sector growth and identifies full-stack platforms, defense robotics, and data-rich scaled operators as likely winners. YC's Jon Xu separately argued robotics is in its AWS-early-customer phase, with infrastructure and training tools as the near-term opportunity.
Why it matters
This flips the 'humanoid bubble' narrative into an under-investment thesis with concrete round-count data. The implication: generalist VCs have been pricing robotics off software comparables and under-allocating to hardware-heavy companies with real operating leverage. Expect capital to shift toward picks-and-shovels plays — simulation, teleop data, actuator supply, fleet orchestration — the layer Xu explicitly flagged and that this week's Skild/Symmetry acquisition and Accenture Robot Brain deployments exemplify.
A Chinese-industry read is that the 'under-investment' is a Western-only problem — Chinese startups are well-capitalized via state funds and corporate LPs, producing the 90% shipment dominance. Skeptics note 42 vs 745 may reflect rational capital allocation: software scales at zero marginal cost, and the 50x projection assumes a ChatGPT-moment Xu himself says hasn't happened.
X Square Robot raised nearly RMB 2 billion (~$292.8M) in Series B co-led by Xiaomi and HongShan, becoming the only Chinese embodied intelligence company with backing from all four of Alibaba, Meituan, ByteDance, and Xiaomi. Capital goes to the WALL-A multimodal foundation model and commercialization via partners like 58 Daojia for home cleaning services. Wall-B, trained on 100+ household data, bundles into home-cleaning robots launching late May.
Why it matters
The investor roster deputizes X Square as the preferred consumer-deployment partner for Chinese super-app distribution channels (Meituan, Taobao, Douyin) and Xiaomi's IoT platform. The Wall-B home-chores pivot is the first major Chinese answer to the 'humanoids are stuck on demos' critique, but 100 households is tiny relative to data scales required for reliable unstructured home manipulation.
Bulls see a platform play — X Square becomes the consumer-humanoid layer on top of China's super-app distribution. Ecovacs, Roborock, and Dreame will not concede the home category without competing directly.
Reliable Robotics closed a $160M Series round led by Nimble Partners to scale production of its Reliable Autonomy System, an FAA-certification-path aircraft automation stack. The company reports 200+ system commitments from commercial and military customers and active deployments under US Air Force and Department of Transportation contracts. Investor roster includes Boeing-affiliated AE Ventures and RTX Ventures.
Why it matters
A $160M round for autonomous aviation with traditional aerospace prime and tier-1 VC participation is a meaningful institutional signal that the certification pathway is credible enough for mainstream aerospace to underwrite. For the autonomy sector broadly, Reliable is the cleanest data point on the question 'can any autonomous system clear FAA Part 23/25 pathways?' — the answer matters for everything from cargo drones to eVTOL to autonomous trucking's NHTSA equivalents.
The 200+ 'commitments' number deserves the standard discount applied to LOIs in aerospace — actual firm orders are the metric to watch. Still, the investor mix (Boeing-adjacent + RTX-adjacent + generalist VC) is the first time an autonomous aircraft company has assembled that combination.
Mariana Minerals deployed Spot for safety monitoring at Copper One, a fully operating Utah copper mine, integrating it directly into MarianaOS as an autonomous mobile sensor platform for continuous thermal, acoustic, and visual monitoring in place of periodic manual inspections.
Why it matters
The Spot-as-sensor-node-inside-a-vertical-software-stack model (MarianaOS) parallels Anduril/Lattice, Skild/Symmetry, and Accenture's Robot Brain — all from this week — confirming that integration into vertical software stacks is the dominant commercial deployment pattern. Mining copper demand projections (US demand ~doubling by 2035 on AI infra + electrification) make inspection automation strategic, not just economic.
This is the durable niche for Spot-class quadrupeds: hazardous industrial inspection where humanoid form factors are unnecessary. The Iberdrola wind substation and LEBO wind-turbine maintenance deployments (April 20) fit the same quadruped-for-inspection pattern.
Qualcomm launched the Arduino Ventuno Q, a sub-$300 SBC with a Dragonwing IQ8 (8-core ARM CPU, Adreno GPU, 40 TOPS Hexagon Tensor NPU) targeting robotics OEMs and developers. Pricing and form factor position it as a direct alternative to NVIDIA's Jetson line, with Arduino-branded distribution into hobbyist, maker, and prototype markets.
Why it matters
A $300 40-TOPS board changes the BOM math on prototype and sub-$5K commercial units where Jetson Orin NX was the default. Combined with this week's DeepX/Hyundai DX-M2 (yesterday) and Google's TPUv8 split (yesterday), the NVIDIA-alternative coalition now spans data center to hobby-tier — consistent with the NPU-over-GPU edge inference trend covered April 20.
Software maturity remains the hard part: CUDA, TensorRT, and Isaac form a moat Qualcomm has historically struggled to match outside mobile. The question is whether a credible ROS 2 / MoveIt / VLA inference stack on Hexagon ships within 12 months.
Taiwan-based Artilux unveiled Inception, a hybrid optoelectronic AI compute architecture replacing conventional ALUs with optoelectronic neurons performing GEMM via analog photonic dot products. The first-generation core targets 12,000+ TOPS (INT8) at ~50mW/mm² on mature CMOS nodes without active cooling. Architecture published in APL Machine Learning with co-authors including Richard Soref and Haisheng Rong.
Why it matters
Photonic/optoelectronic inference has been research-grade for a decade; the distinguishing claim here is 'mature CMOS + no active cooling + 50mW/mm²' — the combination that, if real, enables edge-robot inference densities unreachable with digital silicon. For humanoids and on-board VLA inference where the thermal envelope is the binding constraint, this is the kind of architectural lever that would be a step-change rather than a percentage improvement.
Skepticism is warranted: photonic compute has a long history of spec-sheet TOPS not translating to real workload performance due to ADC/DAC overhead and analog noise. Watch for independent benchmarks on actual LLM/VLA inference rather than GEMM microbenchmarks, and for disclosed customers before treating this as deployable.
San Francisco startup Humble emerged from stealth with $24M seed led by Eclipse Ventures to build the Humble Hauler — a fully autonomous, cabless, battery-electric Class 8-equivalent truck designed for dock-to-dock intermodal freight. Architecture is built on vision-language-action models rather than rule-based systems; removing the cab delivers 360° sensor coverage and higher cargo/energy efficiency. Founder Eyal Cohen previously worked on Uber ATG, Waabi, and Apple autonomy programs; first prototype completed in ~6 months.
Why it matters
Humble is the clearest counter-architecture to the Aurora/Kodiak hub-and-handoff model that currently defines US autonomous trucking. By designing the truck around the autonomy stack (no cab, no driver ergonomics constraints) and adopting a VLA approach rather than the classical perception/planning/control stack, Humble targets the 30–50% logistics efficiency gain that conventional retrofits cannot reach. The 6-month prototype timeline is also a proof-point for how fast VLA-centric teams can iterate versus sensor-fusion incumbents.
Bulls note that cabless designs bypass the federal regulatory debate about driver monitoring entirely — the vehicle is a robot, not a car. Bears counter that dock operators and brokers are deeply human-workflow-dependent and that getting cabless trucks accepted at loading docks will be a multi-year institutional fight. Also worth comparing against Aurora's 200+ truck fleet target by end-2026 (covered April 19) — Humble has the cleaner architecture, Aurora has the deployments.
New detail on the Nissan/Wayve ProPilot story covered April 20: the system is now confirmed on the Ariya EV specifically, targeting near-Level-4 with the same 11-camera/5-radar/LiDAR stack, and is slated for production on the next-generation Elgrand MPV in 2027, Japan first with US rollout to follow.
Why it matters
The explicit 2027 production date on a volume OEM platform is the new information. It directly counters the BMW/Mercedes Level-3 retreat covered April 21 — Wayve's end-to-end learned architecture getting a firm production schedule at the same moment traditional ADAS programs collapse is a meaningful divergence. Tokyo's driving conditions (tight lanes, heavy pedestrian density, ambiguous merging) are a genuine stress test if the 6-week culture-learning claim survives independent review.
Komatsu commissioned its 1,000th ultra-class autonomous haul truck running the FrontRunner Autonomous Haulage System, with the milestone unit a 930E-5AT electric-drive truck (290-tonne payload) deployed at Barrick's Nevada Gold Mines. Cumulative FrontRunner fleet has moved 11.5B+ tonnes of material over ~18 years of commercial operation.
Why it matters
A useful counterweight to the robotaxi and cabless-truck hype: commercial heavy-haul autonomy has been operating at scale for nearly two decades with measured reliability data. The 1,000-unit mark plus the shift to electric-drive autonomous trucks is a real case study in what 'fleet operations' actually look like — the kind of track record Aurora, Kodiak, Humble, and the robotaxi operators are all attempting to build up to.
Mining autonomy succeeds because the operating domain is bounded (private roads, fixed routes, no pedestrians, cooperative operators). Open-road autonomous trucking is a fundamentally different problem, but the financial model — labor displacement, 24/7 utilization, fuel/energy optimization — is analogous and proven.
China's MIIT closed the public comment period April 13 on a 62-page proposed nationwide mandatory safety standard covering SAE Level 3 and Level 4 autonomous vehicles, spanning technical requirements, safety assurance, and testing methods aligned with pending UN regulations. Goldman Sachs projects China robotaxi fleet growth from 5k (2025) to 14k (2026), with robotrucks reaching 760K by 2035.
Why it matters
China is on track to be the first major market with nationwide mandatory AV safety rules at Level 3/4 — regulatory clarity being built in parallel with rapid commercial scaling, unlike the US. The 195% YoY robotaxi fleet growth projection and Geely's purpose-built robotaxi (2027 mass production, announced at Beijing Auto Show) are the commercial evidence behind the regulatory urgency. If China's rules become a global reference — as EU GDPR did for privacy — US AV companies may end up designing to Chinese specs for international optionality.
Type-approval regimes (China, EU) favor incumbents who can afford certification overhead; self-certification regimes (US) favor fast iteration. NHTSA federal equivalents remain absent.
Humanoid enterprise pilots broaden past the Siemens anchor Within 72 hours of the Siemens/Humanoid/NVIDIA Erlangen trial, Accenture+SAP+Vodafone ran a Duisburg warehouse pilot, UBTECH pushed Walker S into ROSSMANN logistics via a DACH distributor, and AgiBot announced its 10,000th unit shipped. The deployment surface is widening from single-site demos to multi-OEM, multi-customer pilots integrated with SAP EWM and Siemens Xcelerator.
Bessemer reframes robotics as under-invested, not bubbly Only 42 robotics firms raised >$30M in five years vs 745 software companies, even as 90% of humanoids ship from China and Bessemer forecasts 50x growth. This narrative — echoed by YC's Jon Xu calling robotics pre-GPT-moment — is the opposite of the 'humanoid bubble' critique and will likely pull more generalist VC capital into the infrastructure/tools layer rather than hardware OEMs.
Edge AI silicon stack fragments further away from NVIDIA Qualcomm's Arduino Ventuno Q at sub-$300 (40 TOPS Hexagon NPU), DeepX/Hyundai's DX-M2 on Samsung 2nm, Arteris+MIPS RISC-V SoC partnership, and Artilux's optoelectronic Inception architecture all hit the same week. Combined with yesterday's Google TPUv8 split (Broadcom/MediaTek/Marvell) and Tesla's AI6 dual-foundry strategy, the NVIDIA-alternative coalition is now visible at every tier from data center to sub-$1k hobby.
Purpose-built robot form factors displace retrofits Humble's cabless autonomous electric truck ($24M seed, VLA-based), Geely's purpose-built robotaxi prototype for 2027, Bosch+ECOVACS's built-in cabinet vacuum, and A&K's airport mobility pod ($8M CAD) all reject retrofit designs. The architectural bet: sensor coverage, cargo/utility density, and serviceability improve faster when the form factor is designed around the autonomy stack rather than inherited from a human-operator vehicle.
Component supply chain is the real moat as humanoids scale Schaeffler's Hermes-winning actuator platform (20% smaller, 500g lighter), the Schaeffler-VinDynamics MoU (second this week), Melexis+OYMotion's Tactaxis tactile modules going to production, and TI's public framing of GaN power stages and deterministic MCUs as bottlenecks all point to the same reality: with component order books full through 2027, whoever locks in actuators, tactile sensors, and MCUs wins deployment slots, not whoever wins benchmarks.
What to Expect
2026-04-24—Hannover Messe 2026 closes — watch for summary announcements on humanoid deployment MoUs and industrial AI integrations.
2026-04-27—Dreame four-day US launch event in San Francisco — hypercar, humanoid, and consumer electronics category expansion from the Chinese vacuum leader.
2026-05-15—General Compute's ASIC-first inference cloud goes GA — a data point on the viability of non-GPU inference economics for agent workloads.
2026-05-28—Texas SB 2807 autonomous vehicle authorization law takes effect; Waymo, Tesla, and Einride must register under the new framework.
2026-08-22—Second World Humanoid Robot Games, Beijing (Aug 22–26) — 100m race goes fully autonomous, real-world scenario events replace simulated venues.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
767
📖
Read in full
Every article opened, read, and evaluated
177
⭐
Published today
Ranked by importance and verified across sources
22
— The Robot Beat
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste