AGV and AMR Battery Packs: Sizing for Continuous Operation

Most AGV battery specifications are wrong in the same direction: too large. The reasoning behind them is borrowed from electric vehicles — work out the runtime needed for a full shift, add a safety margin, specify that capacity. On a vehicle that is charged once a day and driven until it is parked, that logic holds. On a fleet that runs around the clock and charges in the gaps between tasks, it produces a pack that is heavier, more expensive and no more productive than one two-thirds its size.

The reason is structural. An AGV does not have a runtime requirement in the vehicle sense. It has a task chain and a charging window, and the pack only has to carry the fleet from one window to the next. This page sets out how to size a pack for that model, why three classes of mobile robot need three different discharge ratings, and what the BMS accuracy figure actually does to fleet throughput. It draws on the packs we build for wheeled AGV and AMR platforms, quadruped robots and bipedal platforms — the robot and AGV battery pack range.

Key takeaways

An AGV pack is sized against the longest task chain plus the charging window, not against a shift-length runtime target. Sizing by runtime typically over-specifies capacity by a wide margin.
Discharge rating follows the robot’s kinematics, not its weight. Our three platform classes are rated 1C, 2C and 5C respectively for exactly this reason.
State-of-charge accuracy is a scheduling input. A pack that reports SOC to ±2% lets the fleet manager dispatch aggressively; a pack that drifts forces conservative dispatch and idle capacity.
An AGV chassis has no ram airflow. Thermal design has to work in still air, and opportunity charging means the pack spends more of its life charging than a vehicle pack does.
In human-collaborative environments, thermal runaway is a design prohibition rather than a probability to be managed. That, not energy density, is the real argument for semi-solid-state chemistry here.
Mounting position — bottom, back or belly — determines centre of gravity, vibration spectrum and ingress exposure at the same time. It is a structural decision, not a packaging detail.

Why an AGV pack is not sized like an e-bike pack

The e-bike formula is familiar: average assist power multiplied by ride duration, with a margin for hills and cold. It works because the charge cycle is a daily event and the vehicle is idle while it happens.

A fleet AGV inverts both assumptions. It is idle in short windows, several times a day, at a charger it drives to itself. The energy it needs is therefore:

the average traction and payload power, multiplied by
the longest uninterrupted task chain the dispatch system will assign, minus
whatever the charging window can return at the charger’s rate.

That third term is the one buyers forget. If a 45-minute window at 1C returns most of a shift’s consumption, the pack does not need to hold a shift’s energy. It needs to hold the deficit between two windows, plus enough buffer to survive a missed one. A pack sized on the deficit is smaller, lighter and cheaper, and it charges at a higher average state of charge — which is generally kinder to cycle life than deep cycling a large pack.

There is a counter-argument worth stating: a smaller pack means more charging events, and every charging event is a docking manoeuvre the fleet has to schedule. Below a certain fleet size, the coordination cost exceeds the hardware saving. The honest answer is that the optimum depends on charger count and dispatch logic as much as on the pack, which is why the sizing conversation has to include the charger layout rather than starting from a kWh figure.

Three robot classes, three discharge ratings

Buyers often assume discharge rating scales with pack size. It does not — it scales with kinematics. The three platform classes we build for differ by a factor of five on continuous rating relative to capacity, and the spread comes entirely from how the machine moves.

Platform	Usable energy	Continuous discharge	C-rate (charge / discharge)	Typical mount
Bipedal / humanoid	0.85 – 1.4 kWh	40 – 80 A	1C / 1C	Back-mounted
Wheeled AGV / AMR	1.7 – 4.5 kWh	60 – 150 A	1C / 2C	Bottom-mounted
Quadruped / robotic dog	0.45 – 1.0 kWh	50 – 100 A	1C / 5C	Belly-mounted

The pattern in that table is worth reading carefully, because it inverts the intuition that bigger platforms are harder on the battery.

The quadruped has the smallest pack and the highest discharge rating. A legged robot’s power profile is spiky: a walking gait draws moderate current, but a jump, a recovery step or a sprint demands several times the average for a fraction of a second. The duty cycle is low; the peak is not. A pack that cannot deliver 5C will either trip its protection or brown out the actuators at exactly the moment the machine needs torque to stay upright. Note also the density figure on this class — 150 to 200 Wh/kg — matters more here than on any other platform, because every kilogram of pack is a kilogram the legs have to accelerate and arrest on every step.

The bipedal platform needs the highest energy density of the three, and the gentlest C-rate. Humanoid gait is continuous rather than impulsive, so peak-to-average is moderate and 1C discharge is adequate. But balance is a control problem with a mass term: a heavier torso requires more corrective torque, which draws more current, which needs a bigger pack. The datasheets for this class sit at 160 to 210 Wh/kg precisely because density breaks that loop. The gain is not cosmetic: at this scale, mass saved in the pack comes straight off the inertia the controller has to fight on every corrective step.

The wheeled AGV is the opposite trade: modest C-rate, large energy. Traction on a flat floor is predictable, and 2C discharge is enough headroom for acceleration, lifts and payload. What this platform consumes is time, not peak power. A warehouse AMR running sixteen hours with opportunity charging wants the largest usable energy of the three — 1.7 to 4.5 kWh in our range — because every extra hour of autonomy is an hour of dispatch flexibility the fleet scheduler can spend.

The general rule that falls out: size the C-rate from the peak-to-average power ratio of the motion, and size the energy from the task chain. Buyers who specify both from the same number usually get one of them wrong. The electrical relationships behind these figures are the same ones covered in how to size a battery pack, applied to a duty cycle that repeats many times a day rather than once.

SOC accuracy is a scheduling input, not a gauge reading

On a consumer device, state-of-charge error is an annoyance. On a fleet, it is a throughput parameter, and the mechanism is direct. The dispatch system decides which robot takes which task based on how much energy each unit has left. If the reported SOC is optimistic by five percentage points, the system assigns a task the robot cannot finish, the robot stops in an aisle, and the aisle is blocked until someone moves it. If the reported SOC is pessimistic, the system routes the robot to a charger early, and a charger slot that could have served another unit is occupied.

Both errors cost capacity, and neither shows up in a datasheet’s headline capacity figure. This is why the packs we build for this segment publish BMS accuracy separately — SOC to ±2% and SOH to ±2% across the range. Two percent sounds like precision for its own sake until you convert it into dispatch behaviour: at ±2%, a scheduler can plan a task chain to within a small buffer; at ±8%, it has to plan for the worst case and effectively derates the whole fleet.

Getting to ±2% is an algorithm problem more than a sensor problem. Coulomb counting alone drifts, because the current sensor has offset and the offset integrates. A pack that holds accuracy across a month of operation has to correct that drift against something — open-circuit voltage at rest, a full-charge anchor, or a model-based estimator that accounts for temperature and ageing. What matters for procurement is not the method but the evidence: ask for drift measured over repeated partial cycles, not a single full-cycle bench result. The parameters worth checking are the ones set out in BMS specifications explained.

State of health carries the same argument one level up. A fleet manager deciding when to retire a pack, or how many spare packs to hold, is consuming an SOH estimate. If that estimate is slow to react, packs stay in service past the point where their internal resistance has begun to affect dispatch reliability. Publishing SOH to ±2% is a commitment that the retirement decision can be made from the data rather than from a calendar.

Heat in a sealed chassis

A vehicle pack sits in moving air. An AGV pack sits in a sheet-metal box a few centimetres off the floor, in a warehouse held at ambient, with no ram airflow at any point in its duty cycle. Every watt of loss has to leave by conduction into the frame or by natural convection inside a sealed volume. That changes the thermal problem from “manage the peak” to “keep the average down”.

Two heat sources matter, and they peak at different times.

Discharge heat scales with current squared, so it is dominated by acceleration, lifts and gradient work. On a wheeled AGV this is a short, repeatable spike; on a quadruped it is a very short, very large one. Because the peaks are brief, the thermal mass of the pack absorbs them, and the design question is whether the cell-to-cell temperature spread stays acceptable after a hundred repetitions — not whether one spike is survivable.

Charge heat is the one that surprises people. Opportunity charging means the pack spends a larger fraction of its calendar life on charge than a vehicle pack does, and charging at 1C in still air produces a steady, sustained temperature rise rather than a spike. A pack that passes a discharge thermal test can still accumulate heat across four charging events in a shift if the charger and the BMS are not coordinating. This is a system question: charger current, BMS taper behaviour and the physical mounting all contribute, and none of them can be optimised in isolation.

Semi-solid-state chemistry helps here, but the reason is often misstated. The gain is not primarily a higher continuous rating; it is a wider safety margin at the same temperature, which means the BMS can allow a higher charge current at a given thermal headroom without the same consequence if the margin is consumed. For a pack that charges four times a shift in still air, that is worth more than a few extra Wh/kg. The broader state of the chemistry is covered in semi-solid-state commercialisation progress.

Safety: human-collaborative environments change the criterion

In a fenced automated cell, a thermal event is an asset loss. In a warehouse aisle shared with pickers, it is a different category of event entirely. That single change of context is the strongest technical argument for paying the density and cost premium of semi-solid-state cells in this segment, and it is the argument buyers most often leave out of the specification.

The consequence for design is that thermal runaway stops being a probability to be managed and becomes a case that must not occur by design. In practice that means the safety case is built on three things: cell chemistry with a wider tolerance to abuse, mechanical protection of the cell stack against impact from a dropped load or a collision, and a BMS whose protection thresholds are set to intervene before the chemistry is stressed rather than after. None of those are exotic; all three have to be specified together, because a robust cell with a badly tuned BMS is not a safer pack.

It also changes how the pack should be tested. A drop test on the enclosure and a vibration sweep are table stakes. What a safety reviewer for a collaborative deployment will want to see is evidence of behaviour under a single-cell fault — whether the pack contains the event or propagates it. That is a structural question, answered by spacing, barriers and venting, and it needs to be decided at the mechanical design stage rather than validated at the end.

Mounting position decides three things at once

The mounting convention in the table above is not decoration. Each position carries a different combination of structural and environmental load, and moving a pack from one to another is a redesign rather than a relocation.

Bottom-mounted (wheeled AGV) puts the mass where it lowers the centre of gravity, which improves stability under braking and cornering. The cost is exposure: floor-level splash, cleaning water, dust drawn up by the drive, and the highest vibration amplitude in the frame because the pack sits directly over the drive path. The enclosure and the isolation mounts carry that duty.

Back-mounted (bipedal) keeps the mass near the torso’s existing centre, which minimises the inertial disturbance to the gait controller. The constraint is that the pack now moves through the full vertical excursion of the torso on every step, so the mounting interface sees a high-cycle fatigue load rather than a random vibration load. Fastener retention and harness strain relief matter more here than impact protection.

Belly-mounted (quadruped) is the most punishing in peak terms. Ground clearance is limited, so the pack is close to impacts, and every landing from a jump transfers through the mounting points. A pack on a legged platform also experiences load reversals that a wheeled platform never produces, which makes the mounting design a fatigue problem with a very high cycle count.

What buyers should ask for is the vibration and shock spectrum the pack was qualified against, not just an IP rating. An IP65 enclosure on a bracket that fatigues in six months is not a solution. The compliance route for the finished machine is a separate topic, covered in shipping and export compliance.

Charging strategy is part of the pack specification

Three charging models are in common use, and each implies a different pack.

Opportunity charging — the robot docks whenever idle. The pack must accept high charge current without excessive heat, and its BMS has to handle frequent partial cycles without losing SOC accuracy. This is the model the C-rate figures above are written for.
Swap or exchange — a depleted pack is exchanged for a charged one. Here the constraint moves to the connector: cycle life of the contact system, and a mechanical interface that survives thousands of engagements without resistance drift.
Overnight charging — one long charge, deep cycles. Capacity per pack must cover the full shift, and cycle life at depth of discharge becomes the dominant lifetime parameter.

The mismatch to watch for is a pack designed for overnight charging deployed in an opportunity-charging fleet, or the reverse. The first will heat and lose SOC accuracy under repeated partial cycles it was never tuned for; the second carries more capacity than the duty cycle needs and gives up the weight advantage. Because the BMS configuration and the charger have to agree on current limits and taper behaviour, the protocol side matters as much as the electrical side — the message set defined in the BMS specification is where that agreement lives.

Five specification mistakes that cost real money

Sizing capacity from a shift-length runtime target. Over-specifies the pack, adds weight the drivetrain then has to carry, and raises unit cost with no throughput gain.
Ignoring the charging window’s return rate. The capacity requirement is a deficit calculation; without the charger’s current and the window length, the number is guesswork.
Accepting a wide SOC error. It gets converted into conservative dispatch, which is paid for in extra packs across the fleet rather than in a visible line item.
Specifying a BMS without SOH or without a documented protocol. Fleet management needs a health estimate to plan retirement, and integration needs a defined message set. A protection-only board leaves both to guesswork.
Qualifying the pack on a bench instead of on the mounting. A vibration sweep of a loose pack proves nothing about a bracket that resonates at the drive frequency.

A sizing worksheet

Nine inputs settle the design. They mirror the general pack specification but with the duty-cycle terms made explicit, which is where AGV projects usually diverge from vehicle ones.

Average traction and payload power, measured over a real task, not estimated from motor ratings.
Peak-to-average power ratio of the motion — this sets the C-rate, and for legged platforms it can exceed five.
Longest uninterrupted task chain the dispatcher will assign.
Charging window length and charger current, converted into energy returned per window.
Nominal voltage and acceptable window, matched to the drive and the charger.
Envelope and mounting plane, with the centre-of-gravity constraint stated.
Vibration and shock spectrum at the mounting points, plus IP requirement.
Communication protocol and which SOC, SOH and fault frames the dispatcher consumes.
Annual fleet build rate and expected service life, which together set the cycle-life target.

If item 2 is unknown, the pack will be oversized on discharge rating or undersized on peak — a mistake that only appears under a specific manoeuvre. If item 4 is unknown, capacity cannot be settled at all. Those two are worth resolving before a quotation is requested; the rest can follow during engineering. The electrical configuration arithmetic is in lithium cell specifications explained, and the chemistry trade-offs in LFP vs NMC vs sodium-ion.

Frequently asked questions

Is semi-solid-state necessary, or is LFP sufficient for an AGV?

It depends on where the machine operates and what the density requirement is. A wheeled AGV in a fenced cell with generous space for the pack can be served well by LFP, which is cheaper and has a long cycle life. Semi-solid-state earns its premium in three situations: legged platforms where every kilogram matters, collaborative environments where the safety case is strict, and envelopes where the required energy simply does not fit with a lower-density chemistry. It is a constraint answer, not a default.

Can one pack serve several robot models in a fleet?

Usually not without compromise, because the mounting, the envelope and the peak power differ. What can be shared is the cell format, the BMS platform and the protocol — which is where the engineering cost actually sits. Standardising at the platform level while varying capacity and mounting per model is the common middle path, and it is worth designing in at the start rather than retrofitting.

What voltage should an AGV pack use?

48 V, 60 V and 72 V cover most platforms, and all three are configurable within the same product family so the drivetrain can be chosen on its own merits. The voltage decision is really a decision about current: for a given power, a higher voltage means lower current, thinner busbars and less resistive heating — which matters more in a sealed chassis than in open air. The trade-offs are set out in 48V vs 72V battery voltage.

How long does a custom robot pack take?

A standard platform adapted to a new mounting or capacity is fast, because the cell, BMS and enclosure are established. A pack built around a new envelope follows the normal development path — feasibility, BMS and protocol, mechanical design, then sample build and bench testing, typically around two weeks for the sample stage. The sequence and the commercial terms are on the custom pack OEM process page, and the engineering decisions behind each stage are set out in turnkey custom battery pack development.

Do you supply the charger as well?

Charger and pack have to be matched on current limit, taper behaviour and, where the machine controls charging, on protocol. We specify the charging window against the pack’s acceptance rate and work from the charger’s actual output rather than its nameplate. Where the customer already has a charger fleet, that becomes a fixed constraint the pack design works around — which is a legitimate outcome, but it has to be known at the start.

Start a robot pack project

Send the platform type, the mounting plane, the task chain length and the charging window. Those four answers are enough to give a first view on capacity and C-rate, and to say whether the requirement fits a standard platform family or needs a new envelope.

Contact DIDIDEDE, or review the robot and AGV battery pack range first.