Read more about this project in my Thesis .

My time as an MEng student with the MIT Biomimetics lab has been very focused on understanding the design and modelling of battery packs for legged robots. The implementation of batteries in highly dynamic robots is distinctly different from traditional battery use cases such as in electric vehicles. When the goal of your robotic platform is highly dynamic and explosive movements, the main focus of your energy system revolves around achieving high power density. This is similar but also very different from the goal of electric vehicles which primarily chases after high energy density.

Project Motivations

Many of us have seen what hydraulically actuated robotic systems are capable of because of Boston Dynamics. Despite hydraulic valves and pumps being extremely inefficient, the immense instantaneous power output of such hydraulic systems are extremely impressive. Atlas is a clear example of this. Being able to store energy in both elastic diaphrams (I am assuming atlas has this) and being able to divert all hydraulic power to just a few joints is an important benefit to such systems. Power can be sourced from alternate sources like diaphrams (again, this is an assumption) during short durations instead of just relying on your onboard battery pack.

It is clear that the field of robotics is trending towards traditional motor based actuators with the new electric version of Atlas from Boston Dynamics as well as the plethora of electric robots currently being sold by Unitree. The benefits are clear: no hydraulic maintenence, much higher actuator efficiency, much higher actuator bandwidth, and no need for specialized research, manufacturing techniques, and designs for hydraulic actuators. Driving these actuators also requires minimal hardware and can be controlled by a microcontroller.

There is one drawback to purely electric systems that we have already mentioned however: the fact that any output power of the system must be instantaneously sourced purely by the onboard battery. This is where the application of batteries in highly dynamic robots and pretty much every other use case begin to diverge. Electric vehicles, smart phones, laptops, and backup power banks have driven the cost of battery cells extremely low due to the high demand of batteries in these areas. Extensive research has been done to maximize the energy per unit volume and weight to make these devices last longer and go further. These applications typically discharge the battery quite slowly, at a rate that rarely exceeds 1C of output current.

For the case of the MIT Humanoid, trajectories that involve jumping or backflipping instantaneously demand much more than just 1C assuming your battery is reasonably sized. From both simulation and empirical testing using offboard power, it was clear that the MIT Humanoid’s peak output power is definitely going to be limited by its battery. My goal for this project is to build a very high power density battery with a good battery model to predict battery voltage given power draw.

Mechanical Design

Module structure

When I joined the lab, the humanoid was designed for a Dewalt 60V Flexvolt pack. After some first-order battery sizing calculations, the settled upon battery was much larger in size than the dewalt battery if we wanted to perform flips untethered. I chose a 16s3p configuration of 2170 cells because it worked well with our simulations, but also because it was the maximum number of cells I could feasibly fit in the cross section of the robot. The decided upon module structure was a double braced polycarbonate structure. Despite designing battery packs in the past with only one piece of polycarbonate sheet, the high forces and shock loads I expect this battery to see in a robot testing environment pushed me to double brace even if gluing the cells became more of a challenge.

Battery Pack Mechanical Mounting System

The interface between the battery pack and the robot are two tapered dovetails with a positive locking flexture retention mechanism. The pieces are made of aluminum to save weight and are anodized to prevent galling of the surfaces. This ensures that the inertial forces of the battery under acceleration are reacted completely to the sides of the robot in particularly reinforced areas. The dovetail mechanism is coupled to the frame through an ABS interface material, with the theory being that a 50G acceleration (a complete guess) of the battery mass should cause strain in this material of some non-negligible amount, about 0.1% (also chosen somewhat arbitrarily) to absorb some of the energy. ABS was chosen for this interface.

Sidewall Insulation

To insulate the sides of the battery to protect the exposed cell cans, a layer of kapton, PET plastic sheet, and abrasion resistant gaffer tape was applied.

Electrical Design

The battery cells are interconnected with nickel sheets, cut and bent to fit precisely on and around the cells. The cells in this design are all oriented upwards, and to accumulate voltage we must make spotwelded connections to both the center and edge of the cans. The configuration was chosen to be a 16s3p pack. This was quite difficult to route the nickel sheet since the battery is not 3 cell wide, rather 5 cells wide. I spent a long time thinking about how to interconnect the cells, and ended with the following design. There is one section that wraps around on top and there are two spots that need copper sheet reinforcement to handle the current concentration, but overall this design kept the number of unique parts minimal and limited current concentrations to only two areas.

These nickel sheets are inserted after the carrier PCB is applied to the battery however. The carrier PCB both implements distributed temperature sensing as well as solder points to grab the pack node voltages for both monitoring and balancing.

You can see the carrier installed below. This is one of four PCBS in this battery system.

A piece of laser cut double sided tape holds down the PCB.:

Spotwelding

You can see the all the unique nickel sheets parts below except for the sheet that folds back over the top.

Each connection is spotwelded by hand.

Kapton insulation is applied.

The BMS

The electrical system I designed for this battery pack is split into 4 PCBS. That may sound like a lot, but the carrier board which passes through the nickel interconnect and reads cell voltages and temperatures is one of them, and the breakout board that sits inside the humanoid that mates when the battery is plugged in is another. The other two PCBS are the power board and the segment board which is where most of the complexity in this BMS design lies. In the diagram below which omits the breakout board, you can see the way the segment, power, and carrier board are mechanically connected and what signals pass between them.

The dichotomy between the segment and power board was chosen to make maintenence much easier, offloading as much as possible to the segment board which can be removed without having to desolder the high current path which is how the battery cell block below connects to the power board. The two PCBS are connected with a board to board connector because the segment board is a rigid-flex PCB.

Here is an annotated view of the power board and its features:

The power board IRL:

An annotated view of the segment board. This diagram was made before the segment was fully designed so there are some things missing from the PCB.

The segment board IRL:

The PCBs with the rigid flex interconnect:

The segment, the powerboard, and the carrier board assembled into a battery test module with external connection for programming and debugging:

Fully assembled, the battery pack looks like this:

The BOB

There is also the BOB, or the breakout board, that sits inside the humanoid’s torso and waits for the battery to be plugged in.

Firmware in progress

I am currently writing the firmware. The system is currently capable of checking that all cell voltages are within programmed levels, pack voltage and current are within thresholds, controlling the high side nfet switches, putting the BMS into a low power state pulling only 30uA, and beeping an on board buzzer depending on what state transitions of the system occur.

The BMS state will obey the following state machine:

Cell modeling

Will transcribe from my thesis . Read sections 3 and 4 to learn about how I built a battery model that captured battery impedance as a function of the internal time-dependent dynamics of lithium ion chemistry.