By Jae-Won Kim, CTO and Co-Founder
A 32 kg bipedal robot we evaluated through our early-access program carried 4.2 kg of battery housing that added nothing to locomotion. The rigid aluminum box sat at the torso, protecting the pack from impact and holding its cells under uniform compression. It also raised the center of gravity by 6 cm versus distributing the same energy through the leg frames. That 4.2 kg and 6 cm define the design problem examined here.
Why Keep a Single Pack?
Distributed architecture should be compared with the strongest case for a single pack. Keeping all energy storage in the torso gives the BMS designer a simple topology: one cell string, one set of protection FETs, and one thermal management zone. Balancing and fault isolation are straightforward. The pack can also be removed for service, testing, or swapping. For robots exposed to hard impacts or submersion, certifying one sealed enclosure is much easier than certifying a distributed system with connectors routed through every limb.
The CG case is valid in part. Lower CG generally improves static stability margins in humanoid robots, and a pelvis-mounted central pack keeps mass near the body's geometric center, which supports standing balance. That helps explain why most first-generation research and commercial humanoid platforms adopted centralized packs.
Enclosure Mass Problem
The CG case leaves out housing mass. A conventional 30 to 50 Wh cell pack needs a structural enclosure for impact protection, swelling restraint, and attachment to the robot chassis. That enclosure typically contributes 18 to 25 percent of total pack mass. For a 1.2 kg, 36 Wh pack, this means 210 to 300 grams of aluminum or polycarbonate protecting cells that could instead be self-protecting within a structural composite member.
The cable assembly adds another penalty. A torso-mounted pack needs power and signal cables to the hip joint motors, knee joint motors, and ankle actuators. In a 32 kg humanoid, the main power cable bundle for two legs typically weighs 0.9 to 1.4 kg, depending on gauge and routing. Placing structural cells in the leg frames brings the source closer to its actuators, shortening typical cable runs by 60 to 80 percent and reducing cable mass accordingly.
Distributed Integration: The Mass Math
For the 32 kg robot, the central system has 1.2 kg of cells, 0.26 kg of housing, and 1.1 kg of cables, totaling 2.56 kg. The structural leg-frame system has 1.2 kg of cells with the same chemistry and capacity, no dedicated housing because cells are laminated into the CFC leg tube wall, and 0.35 kg of shortened-run cables, totaling 1.55 kg. That saves 1.01 kg at identical energy capacity, roughly 3.2 percent of total robot mass.
At a robot's payload limit, 3 percent of system mass matters. When a robot is designed around an energy budget in which mass determines endurance, the reduction directly extends operating time before recharge. If the leg-frame structural cells provide 36 Wh and the robot draws 150 W average, distributed architecture yields approximately 24 minutes of additional endurance at equivalent range performance.
How Distributed Cells Affect CG
The claim that distributed cells raise CG is incomplete, though not entirely wrong. Leg-frame cells extend from the thigh to the lower leg rather than sitting at the highest point. In a 1-meter standing-height robot, the average height of this cell mass is typically 35 to 45 cm above ground. A central pelvic pack sits at 55 to 65 cm. In typical configurations, moving cells into the leg frames lowers overall mass-weighted CG by 3 to 6 cm compared with a central pack.
The exception is upper-leg concentration, which occurs when engineers keep cells near the hip actuators to minimize cable runs. If that mounting position places the cells above the pelvis centerline, effective CG can rise. Each platform needs a geometry-specific evaluation. The point is not that distributed cells always lower CG, but that the assumption that distribution means higher CG is wrong for most bipedal configurations.
BMS Topology for Distributed Cells
Distributed architecture does add real complexity. When cells are divided across two or four structural leg segments, the system needs either a centralized BMS with monitoring lines to every segment, which restores much of the cable mass, or a distributed BMS with local cell monitors in each leg communicating with a central supervisor over a low-power bus.
Our approach is hierarchical: each leg segment has a local monitoring IC for per-cell voltage and temperature measurement plus active balancing at segment level. The monitors communicate over CAN at 500 kbps with a central supervisor in the torso. Each segment monitoring IC adds 12 to 18 grams per leg, a small amount relative to the cable mass saved. The CAN interface uses a two-wire shielded cable, substantially lighter than a full power cable bundle.
Fault isolation occurs at segment level. If a cell in the right lower leg segment reports an out-of-bounds voltage, the local FET opens that segment from the string. The supervisor then routes the remaining segments to maintain minimum operating voltage. On reduced power, the robot can complete a controlled shutdown or limp to a charging station. This behavior must be defined in BMS firmware before the platform ships. It cannot be added later without significant software revision.
When a Single Pack Wins
This does not mean every mobile robot should adopt distributed structural cells. High-impact robots, platforms operating in abrasive or chemical environments, and systems requiring field-swappable packs for continuous operation have sound reasons to retain a centralized replaceable pack. Structural cells in leg frames cannot be field-replaced without disassembly. If hot-swap capability is required, that requirement alone rules out distributed structural architecture.
Short leg segments also reduce the benefit because there is less geometric advantage in placing cells throughout the limb. A robot with a 200 mm thigh and 150 mm lower leg has insufficient mounting area for meaningful mass distribution. Below roughly 400 mm effective leg length, the integration complexity does not pay off.
Structural cells offer the strongest advantage on platforms with leg lengths of 500 mm or more, operating cycles measured in hours per charge, and a design priority that favors endurance and payload over serviceability. That profile covers a meaningful part of the logistics and inspection robot market, where we are directing our early-access integration work.