Forem: Pomeloid

What Agibot G2 actually does on a tablet production line

Pomeloid — Thu, 16 Apr 2026 06:51:29 +0000

A productive job on the assembly line is concrete: pick a tablet, navigate a factory floor, insert it into a test fixture with millimeter accuracy, sort the result. Here is what that looks like in practice — and what it reveals about where humanoid dexterity stands today.

The task

Longcheer Technology's facility in Nanchang, China manufactures tablets. On the line where Agibot G2 is deployed, the specific workstation is a Multimedia Integrated Testing (MMIT) station — the step where finished tablets are loaded into test fixtures, run through quality checks, and sorted based on results.

This is a precision task. The tablet has to go into the fixture correctly, which means consistent placement at millimeter-level accuracy, repeated hundreds of times per shift, without deviation. It is not a task that tolerates drift.

The robot's job, step by step:

Pick up a tablet from the incoming stack
Navigate the factory floor to the MMIT station
Place the tablet into the test fixture with the required positional accuracy
Wait for the test result
Sort the unit — pass or fail — to the appropriate output location
Return for the next tablet

That loop runs continuously. At 310 units per hour, the cycle time is approximately 19–20 seconds per operation. Agibot also claims cycle times as fast as 12.97 seconds in high-precision automotive assembly, though the Longcheer deployment runs at the 19–20 second figure for this task.

The hardware doing the work

The Agibot G2 is a wheeled humanoid — it moves on a wheeled base rather than walking legs, which is a deliberate choice for factory floor environments where speed and stability on flat surfaces matter more than stair-climbing. It has 26 degrees of freedom (DoF) in its base configuration, expandable to approximately 50 DoF with optional dexterous hands. The arms are dual 7-DoF force-controlled, with sub-millimeter precision claimed for placement tasks.

Perception is handled by a multimodal sensor stack: LiDAR, RGB-D cameras, and multiple RGB cameras providing 360° situational awareness. The hardware is built to automotive-grade component standards with an IP42 rating for the full machine and IP50 for the arms — relevant for a factory environment where dust and occasional liquid exposure are factors.

Power comes from hot-swappable dual lithium battery packs totaling 1,652 Wh, providing approximately 4 hours of runtime per charge. Hot-swap capability means the robot does not need to stop for charging — a second pack goes in while the first charges, enabling 24/7 continuous operation in principle.

What the performance numbers mean

Agibot and Longcheer report three primary metrics for this deployment:

310 UPH (units per hour). This translates to roughly 3,000 units per shift at an 8-hour shift length, accounting for some overhead. The throughput figure is self-reported by Agibot. No independent verification is publicly available.

99%+ success rate in continuous operation. This is the number that requires the most scrutiny. A 99% success rate at 310 UPH means approximately 3 failures per hour — units that need human intervention, rework, or re-insertion. Over a 24-hour operation period, that is roughly 72 interventions. Whether that is acceptable depends on what happens when the robot fails: does the line stop, or does a human step in and the robot continues? The deployment documentation does not clarify this.

140+ hours of cumulative continuous operation with downtime below 4%. This is the most meaningful reliability indicator in the public data. Downtime below 4% over 140 hours means the robot was non-operational for less than 6 hours across that period. That is a real production number, not a demo number — though 140 hours is approximately 18 shifts, which is a limited window.

36-hour line integration. Agibot claims the robot can be integrated into an existing production line within 36 hours. If accurate, this changes the flexibility calculus for humanoid versus fixed automation: a traditional robotic arm integration typically takes weeks to months, including fixture design, programming, and validation. The 36-hour figure has not been independently verified.

What the Longcheer deployment reveals about dexterity today

The MMIT station task is a useful calibration point for where humanoid manipulation actually stands.

It is a constrained, repetitive task with a well-defined success criterion — the tablet is either in the fixture correctly or it is not. The environment is controlled: factory floor, consistent lighting, known object geometry, predictable fixture position. This is not general manipulation. It is a specific pick-and-place workflow executed at production speed.

The significance is not that the robot can do this — fixed automation has done similar tasks for decades. The significance is that a wheeled humanoid can do this without custom-engineered end effectors or dedicated fixtures redesigned around the robot. The G2 uses its standard arms and hands on an existing production line. That is the flexibility argument in concrete form.

What it does not yet demonstrate:

Handling of non-standard or damaged units (how does the robot respond to a tablet that arrives misaligned?)
Performance across task switching — the 310 UPH figure is for this specific task; switching to a different workstation requires retraining or reprogramming
Long-term mechanical wear on the arms under 24/7 operation at this cycle rate

Agibot is planning to expand to 100 robots at Longcheer by Q3 2026, and states expansion into automotive, semiconductor, and energy sectors. Whether the MMIT station performance generalizes to different workstation geometries and task types is the open question.

What we know / what we don't

Verified (per Agibot and Longcheer public statements):

Task: precision loading/unloading at MMIT stations, tablet pick-place-sort
Throughput: 310 UPH, ~19–20 second cycle time
Success rate: above 99% in continuous operation (self-reported)
Cumulative operation: 140+ hours, downtime below 4%
Hardware: 26 DoF base, dual 7-DoF arms, sub-millimeter precision claimed
Integration time: 36 hours (Agibot claim, unverified independently)
Expansion plan: 100 robots at Longcheer by Q3 2026

Not independently verified:

The 99%+ success rate definition (what counts as a failure? what happens when one occurs?)
The 36-hour integration claim
How performance changes across different task types or line configurations
Long-term wear data beyond the 140-hour window

Sources

After 100 hours: what open-source humanoid hardware lets tinkerers actually experiment with

Pomeloid — Wed, 15 Apr 2026 08:43:28 +0000

Menlo Research built walking humanoid legs in 100 days for under $30,000 in R&D — and open-sourced everything. Here is what that actually unlocks.

The problem open-source hardware solves

Humanoid robots are usually closed systems. You cannot inspect the control stack, swap actuators without a vendor relationship, or modify the locomotion policy without signing an NDA. For researchers and builders who want to study how these machines actually work, that is a hard stop.

Menlo Research's Asimov takes a different position. The mechanical designs, simulation models, and control algorithms are open. The "Here Be Dragons Edition" DIY kit — available for pre-order since Mar 2026 at $15,000 — is priced close to bill-of-materials cost. The company states that low-volume manufacturing should come in under $20,000 for the full body.

The question worth asking: once you have built it, what can you actually do with it that you could not do before?

What you are building

The Asimov v1 robot stands 1.20 m tall, weighs approximately 35 kg, and has 25+2 degrees of freedom (DoF) — the additional two DoF come from articulated toes. It uses a modular architecture: legs, torso, arms, and head snap together via universal motor mounting fixtures, so subsystems can be developed and tested independently.

The leg subsystem costs just over $10,000 all-in — roughly $8,500 for actuators and joint parts, the remainder for batteries and control modules. Most structural parts are compatible with Multi Jet Fusion (MJF) 3D printing, which produces functional parts without custom tooling. Teams without industrial MJF access can use alternatives including casting or standard 3D printing for most components. The knee plate — where stiffness and alignment are most critical — requires CNC machining, but Menlo redesigned it specifically to simplify that requirement.

Assembly takes approximately 100 hours, per Menlo's own account of building the legs subsystem.

What the architecture opens up

Locomotion research without lab access. The ankle design uses a parallel RSU (Revolute–Spherical–Universal) architecture rather than a simple serial joint. This gives two DoF — roll and pitch — with torque sharing between two motors, proximal placement of heavy components, and better backdrivability, meaning the ankle responds more naturally to ground contact forces. The design choices are documented, which means a tinkerer can study why they were made, test modifications, and observe the effects on gait directly.

Sim-to-real experimentation. Menlo built a parallel in-the-loop (PIL) simulator with intentional latency injection to match real hardware behavior. This is the gap where most academic research stalls: a control policy trained in simulation behaves differently on physical hardware because the simulator was too clean. With a documented PIL setup and the physical robot to test against, a builder can iterate on the sim-to-real transfer problem directly — not just read about it.

Gait tuning on your own hardware. Commercial robots ship with locked locomotion stacks. Asimov ships with an open control policy. That means a tinkerer can modify gait parameters, test different walking strategies, and measure the results on physical hardware — not just in simulation. For a researcher interested in locomotion, that is the difference between studying the subject and working on it.

Custom sensor integration. The modular architecture and open mechanical designs make it feasible to add sensors — cameras, force-torque sensors, tactile arrays — without voiding a warranty or waiting for vendor support. Each module has clear interfaces and documented load paths.

Manipulation research on the full stack. Different labs have different priorities. The modular design lets a team working on arm manipulation use just the torso and arms, while a team focused on locomotion works with the legs independently. A tinkerer building the full robot can switch focus without rebuilding the hardware.

What we know / what we don't

Verified:

Kit price: $15,000, close to BOM cost per Menlo Research
Full robot target: approximately $30,000 assembled
Specifications: 1.20 m tall, ~35 kg, 25+2 DoF
Leg subsystem cost: just over $10,000 all-in
Build time for leg subsystem: under 100 days at R&D pace
Manufacturing approach: MJF 3D printing for most structural parts, CNC for knee plate
Ankle design: parallel RSU architecture, documented
Pre-order available since Mar 2026, $499 deposit

Not verified / unknown:

Real-world build time for a tinkerer working alone (100 days was a team R&D effort)
Actuator pricing from Encos (vendor does not publish; requires direct quote)
How many kits have shipped or are in customer hands as of Apr 2026
Long-term part availability and community ecosystem maturity

What to watch

The Asimov kit is early. "Here Be Dragons" is not a product disclaimer to ignore. The questions that will determine whether this becomes a meaningful research platform:

Does a functioning community form around shared modifications, gait improvements, and sensor integrations?
Can builders outside well-resourced labs actually source and assemble the hardware at the stated cost?
How does the open locomotion stack perform compared to proprietary systems on standard benchmarks?

The modular architecture and open software stack are the right foundations. Whether the kit delivers on its research premise depends on what the first builders do with it — and whether they share what they find.

Sources

Why a half-marathon is a better robot benchmark than a warehouse demo

Pomeloid — Tue, 14 Apr 2026 20:33:14 +0000

A 21.1 km road course exposes failure modes that controlled environments never surface. Beijing is about to run the test again.

The demo problem

Watch enough humanoid robot videos and a pattern emerges. The robot walks gracefully across a clean floor. It picks up a box from a table. It waves. The crowd applauds. The founder says something about the future of work.

What the video does not show: the 40 takes before the clip that made it. The controlled temperature. The surface calibration. The team of engineers just out of frame.

A half-marathon fixes all of that. Twenty-one kilometers of public road, shared with thousands of human runners, over two-plus hours of continuous operation, in whatever weather shows up. No cuts. No resets. Either the robot crosses the finish line or it does not.

On Apr 19, 2025, 21 humanoid robots found out which category they fell into. Six finished.

What 2025 actually showed

The inaugural Beijing humanoid robot half-marathon — organized by the Beijing E-Town district, which has positioned itself as a hub for China's robotics sector — was structured as a human-robot co-run. Approximately 12,000 human runners shared the course with the robots. The fastest human finished in 1 hour and 2 minutes.

The fastest robot, Tiangong Ultra (developed by the Beijing Humanoid Robot Innovation Center, also known as X-Humanoid), finished in 2 hours, 40 minutes, and 42 seconds. That is a pace of approximately 7.9 min/km. Tiangong Ultra required three battery swaps during the race and fell once.

Second place went to Noetix Robotics' N2 at 3 hours and 37 minutes. DroidUp's X02 finished third at 4 hours and 50 minutes. Only Tiangong Ultra beat the human cutoff time of 3 hours and 10 minutes.

Fifteen robots did not finish. The failure modes were instructive:

Thermal failure: at least one robot overheated to the point of visibly smoking and had to withdraw
Battery depletion: several robots ran out of charge before completing the course
Balance failure: robots fell on the cambered road surface — a different challenge from flat lab floors
Mechanical breakdown: joints and connectors that perform in short demos failed under sustained load
Navigation errors: some robots drifted, spun in circles, or required physical human assistance to stay on course

Human intervention was common. Support teams steadied robots, changed batteries mid-race, and in some cases physically guided robots through sections of the course.

The completion rate — 6 out of 21, or 29% — is not a failure of the concept. It is the point of the concept.

What the course actually tests

A half-marathon applies four simultaneous stresses that controlled environments routinely avoid.

Thermal management. Actuators and joint motors generate heat. In a 10-minute demo, that heat dissipates. Over two hours of continuous operation, it accumulates. Current humanoid robots are not designed with thermal budgets for multi-hour operation — a constraint that also applies to factory shift work.

Battery endurance. Most humanoid robots in 2025 had runtimes in the 60–120 minute range under moderate load. Running a half-marathon at speed exceeds that budget. The teams that finished either had larger battery packs, mid-race swaps, or both. Tiangong Ultra required three swaps.

Dynamic balance on uneven surfaces. Laboratory floors are flat and consistent. Roads have camber — a slight slope toward the curb for drainage — plus minor surface variations, cracks, and painted markings. Balance controllers tuned on lab floors encounter unexpected inputs on a real road. The robots that fell in 2025 largely fell on these surface variations.

Mechanical wear under sustained load. A robot that performs cleanly for a 5-minute demo has been validated for 5 minutes. The same robot running for 2.5 hours has a different profile. Joints develop play. Connectors loosen. Seals fail. A marathon is a durability test that no demo can replicate.

These four failure modes are not marathon-specific. They apply directly to industrial deployment: an 8-hour factory shift involves sustained thermal load, battery management, variable floor surfaces, and mechanical wear. The half-marathon does not prove a robot is ready for a factory. But it surfaces the gaps that would prevent it.

What is different in 2026

The 2026 Beijing E-Town Humanoid Robot Half-Marathon, scheduled for Apr 19, 2026, is a significantly larger event. More than 300 humanoid robots from over 100 teams are registered — a fivefold increase from 2025's 21 robots. The field includes 76 institutions across 13 Chinese provinces, more than 80 corporate teams, over 20 university teams (a tenfold increase from 2025), and four international teams. Twenty-six distinct robot brands are represented.

The scale change is notable, but the structural change matters more.

In 2026, approximately 38% of competing teams will deploy robots capable of fully autonomous navigation — no remote control, no human joystick guidance. This is a new category. A robot that navigates a half-marathon autonomously has demonstrated something qualitatively different from a robot that is steered by a human operator: it has managed its own path-planning, obstacle response, and course-following over 21 km of real-world terrain.

A full-scale test run involving more than 70 teams — including four international teams — was conducted Apr 11–12, 2026, to stress-test logistics: timing systems, emergency support protocols, battery swap procedures, and complex terrain scenarios. The organizational infrastructure is more serious than 2025.

What to watch next

The 2026 race results will be available after Apr 19. These are the metrics worth tracking:

Completion rate: did it improve significantly from 29%? Any rate above 50% would suggest meaningful progress in endurance engineering
Autonomous navigation completions: how many of the 38% autonomous teams finished without human intervention?
Fastest time: did any robot approach or beat Tiangong Ultra's 2:40? A sub-2:00 would be a significant step
Failure mode distribution: did thermal and battery failures decrease relative to 2025, or did new failure modes emerge at scale?
International team performance: the four non-Chinese teams provide a useful comparison point for where the broader industry sits relative to China's robotics cluster

The race is not a deployment readiness test. But it is the closest thing to a standardized, public, multi-vendor benchmark that the humanoid robotics sector currently has. Every robot that falls on that road is generating data that no lab can produce.

What we know / what we don't

Verified:

2025 completion rate: 6/21 (29%)
2025 winner: Tiangong Ultra, 2:40:42, three battery swaps, one fall
2025 failure modes: thermal, battery, balance, mechanical, navigation
2026 field size: 300+ robots, 100+ teams, 26 brands
2026 autonomous navigation category: ~38% of teams
Full-scale test: conducted Apr 11–12, 2026

Not yet known (as of Apr 14, 2026):

2026 race results (race is Apr 19)
Whether autonomous navigation teams will complete the course
Whether thermal/battery engineering has improved sufficiently