duckbot

Scientific research is results-driven

Sometimes by any means necessary

Duckweed is a small aquatic plant with a very simple structure; just a frond which floats on the water, and some species have roots, and they're quite small with different species ranging from fractions of a millimeter wide to a couple millimeters. Here we're looking at some duckweed in a well plate, and a thermal image of a 96-well plate. As I've learned from working with duckweed scientists, it's these sorts of images that might provide information to answer particular research questions; maybe we're interested in tracking how that duckweed grows over time, under various environmental conditions. Less relevant to answering those sorts of questions is the story of how that duckweed got into that well. Probably, a research assistant was responsible for moving individual fronds into each well by hand. Not only is this sort of manual work time-consuming and tedious, but there's also high stakes; in order to trust the final data, we need to be confident that the right duckweed went into the right wells, which were filled with the right media, which are adequately randomized around the well plates, and so on. This sort of work is currently a necessary bottleneck in getting to some desired result.

Previously: “dull, dirty, and dangerous”

Robots have historically been for mass-production.

They are expensive, tedious to program, and difficult to customize & run.

Automation as a Tool for Creative Exploration

In negotiating these two positions, of fully manual and fully automated approaches, our lab is interested in automation as a tool for creative exploration. For science, that might look something like this: defining an experiment up front in a way that accommodates niche experimental design; being able to compile that experiment to instructions which can be run on a machine-- and maybe this requires physically configuring the machine for your custom application, and finally collecting data, where the operator--the scientist in this case-- can have some role to play at runtime. In addition to, say, scaling up an individual experiment, our research interest here is in how we can provide flexible infrastructure which allows scientists to build, customize, and run machines across a range of applications. It takes significant time, skill, & effort to build a system for a specific task. We want to lower the threshold to automation in a way that allows scientists design and reuse systems without compromising the level of control over the experiment.

Automation as a Tool for Creative Exploration

requires...

Modular, Low-cost, Extensible Hardware

Human-centered Systems for Programming/Running Experiments

This picture of automation requires some infrastructure distinct from that of the mass-automation machines we were looking at. First, niche applications require niche tools, we can't take advantage of high volume sales & production to offset costs. Instead, we need modular, low-cost, and extensible hardware to customize one-off applications. And after we build a machine, we need good ways of actually running it. Control is commonly enacted from an interface built for a specific task. We are interested alternative controls, including writing code which compiles to machine instructions to allow reusability and flexibility in defining experiments, as well as live sensor readings to interactively control the flow of the experiment at runtime. These are two areas of research that our lab has been been actively engaged in.

Automating Growth Assays

define experiment → setup plates → collect data → analyze data

Automating Growth Assays

define experiment → setup plates → collect data → analyze data

 	Experiment_name = "Demo" 
	genotypes = ["Sp7498", "Lm5500", "Lm8627", "Wa7788"]
	media = ["0mM", "25mM", "50mM", "100mM"]
	reps  = 4

Let's start with defining the experiment. As we step through this workflow, I'll highilight snippets of code as they appear in our Jupyter notebooks-- more code than just this is requried to actually run the experiment, but this should convey the level of control we can have over the experiment to suggest alternative use cases, too. Here, we're defining an experiment with 4 duckweed genotypes, 4 media solutions of varying salt molarity, and 4 replicates of each condition. This 4x4x4 experimental design will already requires 64 wells, which approahces a scales tedious to plan out by hand, if not necessarily high-throughput. Our experimental setup will randomize the location of each replicate around the well plates, and the pre-visualization of our setup looks like this for the first plate. (show video) We can see in here how running code which uses this experimental data as input can set up our experiment for us. A couple important notes and limitations as we watch the syringe tool go around here. First, we're using 24-well plates here. We could also design experiments which use 12, or 96 well plates. Size cononstraints from the Jubilee mean we have room for 6 well plates or reservoirs. In our demo experiment, we already have 4 media and 3 24-well plates, meaning we've exceeded this limit. We've designed our software with human-in-the human-in-the-loop actions in mind, meaning duckbot will guide us through physical configurations like swapping media when necessary with simple text prompts. Moreovoer, we can take advantage of the interchangeable bed plates for experiments requiring greater than 5 well plates. And finally, our machine is being operated in the open. Sterility is a concern moving forward, and we can think about differently sized machines that might, say, be able to be operated within a laminar flow hood.

Automating Growth Assays

define experiment → setup plates → collect data → analyze data

With our plates setup, we can take images of each well. Here we see duckbot grabbing its camera and moving to the first well. We can the go down the line to save close-up images of each well. Cutting to the camera's view, we see here duckweed with the frond contours highlighted. This sort of straighforward image analysis can be integrated into experiments on the fly. By identifying fronds, we can differentially manipulate samples depending on, say, frond area, or growth rate by pulling data from previous days. Imaging individual wells provides high quality top-down images. It takes approximately a minute per 24 well plate. Images are saved with consistent file naming conventions which id each well in a human-readable way. We can alternatively take images of all well plates at once by dropping the bed down. We can then segment individual wells from the full image.

Custom Worfklow Automation with Jubilee

Machine Agency
Jan 28, 2023

	m = MachineCommunication(port)
	m.toolChange(tools['media-syringe'])
	m.moveTo(x=media_reservoir['x'], y=media_reservoir['y'])
	m.aspirate(mL=20)
	wellA1 = fetch_well_position(1, "A1") # (plate, well id)
	m.moveTo(x=wellA1['x'], y=wellA1['y'])
	m.dispense(mL=1.5)

	m = MachineCommunication(port)
	m.toolChange(tools['media-syringe'])
	m.moveTo(x=media_reservoir['x'], y=media_reservoir['y'])
	m.aspirate(mL=20)
	wellA1 = fetch_well_position(1, "A1") # (plate, well id)
	m.moveTo(x=wellA1['x'], y=wellA1['y'])
	m.dispense(mL=1.5)

	m = MachineCommunication(port)
	m.toolChange(tools['media-syringe'])
	m.moveTo(x=media_reservoir['x'], y=media_reservoir['y'])
	m.aspirate(mL=20)
	wellA1 = fetch_well_position(1, "A1") # (plate, well id)
	m.moveTo(x=wellA1['x'], y=wellA1['y'])
	m.dispense(mL=1.5)

	m = MachineCommunication(port)
	m.toolChange(tools['media-syringe'])
	m.moveTo(x=media_reservoir['x'], y=media_reservoir['y'])
	m.aspirate(mL=20)
	wellA1 = fetch_well_position(1, "A1") # (plate, well id)
	m.moveTo(x=wellA1['x'], y=wellA1['y'])
	m.dispense(mL=1.5)

	m = MachineCommunication(port)
	m.toolChange(tools['media-syringe'])
	m.moveTo(x=media_reservoir['x'], y=media_reservoir['y'])
	m.aspirate(mL=20)
	wellA1 = fetch_well_position(1, "A1") # (plate, well id)
	m.moveTo(x=wellA1['x'], y=wellA1['y'])
	m.dispense(mL=1.5)

Custom Worfklow Automation with Jubilee

Scientific research is results-driven

Previously: “dull, dirty, and dangerous”

Automation as a Tool for Creative Exploration

Automation as a Tool for Creative Exploration

Modular, Low-cost, Extensible Hardware

Duckbot

Duckbot

Duckbot

Duckbot

Duckbot

Duckbot

Duckbot

Duckbot

Duckbot

Duckbot

Duckbot

Automating Growth Assays

Automating Growth Assays

Automating Growth Assays

Automating Growth Assays

Automating Growth Assays

Automating Growth Assays

Root Imaging

Root Imaging

Root Imaging

So you want to build a Jubilee!

So you want to build a Jubilee!

So you want to build a Jubilee!

Building & Extending a Duckbot

Building & Extending a Duckbot

Custom Worfklow Automation with Jubilee

Pourover Slides!

Growth Curve

Syringe Transfer

	Experiment_name = "Demo"
	genotypes = ["Sp7498", "Lm5500", "Lm8627", "Wa7788"]
	media = ["0mM", "25mM", "50mM", "100mM"]
	reps = 4