**Xiaomi puts its humanoids to the test: 8 hours of sorting auto parts in a factory, anomaly detection included**

Xiaomi puts its humanoids to the test: 8 hours of sorting auto parts on a factory line, anomaly detection included

🔎 Endurance becomes the new standard for the industrial humanoid

On November 3, 2026, two Xiaomi CyberOne 02 humanoids held a sorting station for center console side covers for 8 hours and 5 minutes straight at the Xiaomi Auto plant in Beijing. No downtime, no visible teleoperation: the robots detect anomalies, coordinate with the line's MES, and call a mobile robot (AMR) when a part drops. This is the first public demonstration of a full shift in real conditions — not a 20-minute lab demo. The industry is pivoting: the metric that matters is no longer "does it work?" but "how long can it run without human intervention?"

The essentials

• Xiaomi validates industrial endurance: two CyberOne 02s sorted automotive parts for 8 h 05 on November 3, 2026, with two autonomous recharges and multi-robot coordination (humanoids + AMR).
• The US race answers back: AGCO and Agility Robotics are launching a Digit pilot at the Hesston, Kansas plant for 2027 production — the first American deployment at factory scale.
• The "Physical AI" infrastructure opens up: Serve Robotics is monetizing its stack (connectivity, data curation, catalog) with third parties like Bedrock Robotics — a sign that value is migrating from hardware to platform.

Recommended Tools

The Xiaomi demo: what the 8 hours prove (and what they leave unsaid)

The 5 min 59 s video published by The Paper (澎湃) shows two bipedal CyberOne 02 units and a wheeled-base robot operating at the SPS (Sorting Supply) station of the final assembly shop. The task: grasping console side covers — soft, large parts in random poses — and placing them on the line at the imposed pace. It's a test of bimanual manipulation, anti-interference visual perception, and force/position adaptation, not structured pick-and-place.

What is demonstrated: holding a station through a full shift, autonomous anomaly detection/correction (misoriented part, dropped part), coordination with the line's MES, multi-robot collaboration (the AMR retrieves the fallen part and restocks the bins), energy management via two recharges (at 3 h 02 and 6 h 03 of cumulative work).

What remains opaque: actual throughput (parts/hour), successful picks vs attempts, number of off-camera human interventions, recharging method (battery swap? tether? inductive charging?), hourly operating cost. Xiaomi publishes no quantified KPIs. Lei Jun answers the skeptics: "it's a general-purpose humanoid, the hardware needs no modification to adapt to more stations and scenarios." The CyberOne 01 (21 DoF, 3D reconstruction with 1% accuracy at 8 m, 85 environmental semantics, 45 emotions) remains the public spec sheet — the 02 has no datasheet.

The program in three milestones: from March to November 2026

Electrek places the demo within a structured annual program:

The progression is clear: growing task complexity, growing duration, rising success rate. The video, sped up to 100x, hides the real elapsed time — an 8-hour shift compressed into 6 minutes. Humanoids Daily notes that "eight hours of task does NOT mean eight hours on a single charge." The post describes neither a swap nor a charge; the most likely hypothesis is fast recharging during the line's micro-pauses (changeovers, scheduled maintenance).

AGCO / Agility Robotics: the American response in Kansas

While Xiaomi films in Beijing, AGCO announces on October 20, 2026 a Digit pilot (Agility Robotics) at its Hesston, Kansas plant, targeting production in 2027. Tomorrow's World Today specifies that Digit will work on internal logistics tasks: moving bins, loading/unloading lines, replenishing workstations. Unlike Xiaomi (a pure bipedal humanoid), Digit has a wheeled base for long-distance mobility and a manipulator torso for grasping — an unapologetically hybrid architecture.

The AGCO/Agility partnership marks the first public commitment by a major American manufacturer (agricultural equipment maker, $11.4 B revenue in 2025) to a humanoid deployment at factory scale, not lab scale. Agility raised $150 M in a Series B (March 2026, $1.2 B valuation) to industrialize Digit V4. The timeline: pilot in Q1 2027, scale-up decision in Q3 2027. If it holds up, this will be the first humanoid production contract in the US.

Serve Robotics: The "Physical AI" Stack as a New Business Model

On November 5, 2026 — two days after the Xiaomi demo — Serve Robotics launches Physical AI Services: three products that transform its fleet's infrastructure (2,000+ robots, 40+ cities, 27 hospitals) into a sellable platform.

Ali Kashani (CEO): "Building a robot is only the beginning. Keeping thousands of them connected, learning, and improving requires another layer of technology." Serve, an Uber spin-off (2021), acquired Diligent Robotics (2026) to enter the hospital services market. StreetInsider confirms the three products and the Bedrock customer. The model: stop selling robots, sell the intelligence that keeps them running in production. Xiaomi, Figure, Tesla, Agility will all have to solve this problem — or buy it.

Why Endurance Is Becoming the #1 Criterion

The demos of 2023–2025 (Figure 01 at BMW, Optimus Gen 2 at Tesla, Digit at Amazon) showed isolated tasks: 20–60 minutes, controlled environment, engineers on standby. The 8-hour shift is a game changer for three reasons:

1. Thermal and energy management: a 50–60 kg bipedal humanoid consumes 300–500 W during continuous manipulation. Sustaining 8 hours requires either a 4–5 kWh battery (heavy, bulky) or battery swap/fast charging integrated into the production cycle. Xiaomi did two recharges — the method isn't public, but the integration into the line's rhythm is.

2. Perception/control drift: over 8 hours, lighting changes, parts wear out, tolerances drift. Autonomous anomaly detection (Xiaomi: "detect and correct anomalies autonomously") requires a perception-action loop that is robust to drift, not just a frozen RL policy.

3. Multi-agent coordination: the AMR that picks up the dropped part and restocks the bins proves the humanoid isn't an island — it's a node in a cyber-physical system. This is what Figure Helix-02: humanoid robots complete full 8-hour shifts in a factory — without human intervention also illustrates on the US side: endurance emerges from orchestration, not from the robot alone.

The invisible software stack: what really keeps the shift running

Behind the 8 h 05, there's a software stack that Xiaomi doesn't detail but whose building blocks can be inferred:

• Real-time 3D perception: scene reconstruction at >30 Hz, anomaly detection (missing part, aberrant orientation) without retraining — probable RGB-D fusion + visual foundation models like DINOv2/SAM 2.
• Whole-body MPC control: model-predictive optimization over 21+ DoF (bimanual + base) with collision, effort, and line-rate constraints. The CyberOne 01 is announced at 21 DoF; the 02 likely adds dexterous hands (12-16 DoF/hand).
• MES/ROS 2 orchestration: real-time bridge between the production system (Siemens/GE/Dassault MES) and the ROS 2 fleet. The AMR called in for the dropped part = action triggered via ROS 2 topic /action/recover_part.
• Predictive energy management: SoC/SoH estimation, recharge scheduling during line windows, tether fallback if the swap fails.

NYU releases MABEL, a mobile bimanual robot with 56 degrees of freedom for $8,056 shows that academic research is converging toward the same architectures: whole-body control + bimanual manipulation + mobile base. The difference: Xiaomi is industrializing it at factory scale.

Safety and cybersecurity: the blind spot of the demos

No source mentions ISO 13849 certification (functional safety), IEC 62443 (industrial cybersecurity), or ISO/TS 15066 (collaborative robots). A 60 kg humanoid operating in contact with an automotive production line without a safety cage raises questions that the video does not answer:

• Emergency stop: stopping time < 200 ms? Hardware redundancy?
• Human presence detection: 360° lidar? ToF cameras? Sensor fusion?
• Network segregation: is the ROS 2 DDS on an isolated VLAN? Encryption? Mutual authentication?
• OTA updates: signing, rollback, canary deployment across the fleet?

Vicone finds 28 zero-days in robotics in under 48 hours — LIA accelerates reminds us that the robotic attack surface (ROS 1/2, OPC UA, MQTT, controller firmware) is vast and poorly audited. Xiaomi, like Figure and Tesla, will have to publish compliance artifacts before industrial scale-up. The 8-hour demo is impressive; the security audit is less so — and it is mandatory for production.

Unit Economics: The Number Nobody Publishes

What does an 8-hour shift cost with two CyberOne 02 units + 1 AMR + charging infrastructure + supervision? Low-end estimate based on known public costs (November 2026):

A human operator costs (salary + benefits + training): $25-35/h in China, $40-60/h in the US. The humanoid is not competitive in pure cost terms today. The bet: generality (same hardware, 50+ workstations), flexibility (software reconfiguration, not mechanical), data (every shift enriches the model). Xiaomi is banking on multi-workstation amortization — the CyberOne 01/02 as a platform, not a dedicated machine.

The competitive landscape: three arches, one race

Foxconn presents its first humanoid robots in Europe at VivaTech: the "physical AI closed-loop" stack that is transforming the factory shows that integrators (Foxconn) are becoming players in their own right: they don't just buy, they co-develop the closed perception–action–data loop. The week humanoids went public: Agility SPAC, Unitree in Shanghai, and Tesla converts its plant for Optimus confirms the financialization: access to public capital determines the pace of scaling.

What separates the demo from production: 5 unmet thresholds

1. MTBF > 2,000 h: Industrial Mean Time Between Failures. 8 h = 0.4% of the threshold. No public data on humanoid MTBF.
2. Intervention rate < 0.1%: one line stoppage per 1,000 picks. Xiaomi: rate not published. Electrek: "Xiaomi does not specify the number of sorted parts nor how the robots lasted 8 hours."
3. Reconfiguration < 4 h: changing stations (sorting → screwing → inspection) without a robotics engineer. Lei Jun: "hardware does not need modification" — what about the software?
4. Total cost of ownership < 2x human: threshold for massive adoption. Current estimate: 4-8x.
5. Functional safety certification SIL 2/PL d: mandatory in operator zones. No player has published a certificate.

Common Mistakes

Mistake 1: Confusing a video demo with industrial KPIs

What goes wrong: Taking "8 hours" as proof of maturity without asking about throughput, success rate, interventions, or recharging method.
The solution: Demand the 5 metrics: parts/hour, % of successful picks, human interventions/shift, recharge cycles/shift, cost/effective hour.

Mistake 2: Believing hardware is the bottleneck

What goes wrong: Comparing DoF, motor torque, hand payload. The CyberOne 01 (21 DoF) and 02 do the same demo — the delta is software.
The solution: Evaluate the stack: perception, whole-body control, MES orchestration, energy management, MLOps. Serve Robotics proves it: value is shifting to the platform.

Mistake 3: Ignoring industrial cybersecurity

What goes wrong: Deploying a ROS 2 fleet on the factory network without segmentation, encryption, SBOM, signed updates.
The solution: Demand IEC 62443-4-2, ISO 21434 artifacts, signed SBOMs before any pilot in the operator zone. Vicone finding 28 zero-days is no anecdote.

Mistake 4: Underestimating MES/ERP integration

What goes wrong: Assuming the robot "magically adapts." Xiaomi's coordination (MES + AMR + humanoids) required months of specific engineering.
The solution: Budget 30-50% of total CAPEX for software integration, standardized interfaces (OPC UA FX, ROS 2 DDS Security), end-to-end testing.

Frequently Asked Questions

Will Xiaomi sell the CyberOne 02 to other factories?

No announcement. The current strategy: internal deployment (auto factories, possibly smartphones) to validate the general-purpose platform. Lei Jun insists that "the hardware doesn't need modification to adapt to more workstations" — a sign of platform ambition, but the business model (sale? rental? RaaS?) is not public.

How many CyberOne 02 units exist?

No official figures. The demo shows two in parallel + one AMR. The three-milestone program (March/July/November) suggests a fleet of 5-10 units in continuous iteration in Beijing. No multi-site deployment announced.

Does the CyberOne 02 use RL (reinforcement learning) or classical programming?

Probably hybrid: whole-body MPC for basic locomotion/manipulation, RL policies for adaptive grasping (soft parts, random orientation) and anomaly detection. Xiaomi doesn't publish the architecture. The visual foundations (DINOv2, SAM 2) are near-certain for zero-shot perception.

Why Xiaomi and not a robotics pure-player (Figure, Agility)?

Vertical integration: Xiaomi designs its factories, its lines, its MES, its chips (Pengpai/Surge), its AI models. The humanoid is one more sensor-actuator in a system they control end-to-end. Figure/AGCO have to negotiate the interface with the customer's MES — additional friction.

What is the role of the wheeled AMR in the demo?

It retrieves the fallen part (which the bipeds can't pick up efficiently from the floor) and restocks the feeding bins. Proof that the optimal factory architecture = bipeds (fixed-station manipulation) + AMR (transport/logistics). Digit (Agility) merges the two into a single hybrid robot — a different bet.

✅ Conclusion

Xiaomi's 8 hours aren't an end — they're the first credible data point on humanoid endurance in real production. The race is shifting: 2023-2025 = "it works in demos"; 2026-2027 = "it lasts a full shift"; 2028+ = "it costs less than a human across 50 workstations." Three archetypes are competing: the pure generalist biped (Xiaomi, Figure, Tesla), the logistics wheeled+torso hybrid (Agility, Apptronik), and the "Physical AI" platform that sells intelligence, not bodies (Serve). The winner won't be the prettiest robot — it will be the one that publishes first an MTBF > 2,000 h, a cost/hour < 2× human, and SIL 2 certification. Demos, meanwhile, are free.