—And why it's costing you more than you think.

In 2024, we set out to solve a problem that most robotics teams had learned to live with: residual vibration.

Not because it was unsolvable — the physics and control theory have existed for decades. But because solving it required bridging a gap that the robotics industry had quietly accepted. On one side, affordable robot hardware that's proliferating across every sector. On the other, motion performance that still depends on manual tuning, conservative speed limits, and a lot of patience.

Residual vibration sits right in that gap. And it's more expensive than most teams realize.

The problem hiding in plain sight

Every robotics engineer has seen it. The arm reaches the target position, and instead of stopping cleanly, it shakes. Sometimes for a few milliseconds. Sometimes long enough to watch.

What's actually happening is straightforward physics. A robot arm is not a perfectly rigid machine. It has structural compliance, actuator dynamics, cable routing, and real-world properties that the commanded trajectory doesn't account for. When the joints reach the target, the rest of the system hasn't fully settled. Energy is still stored in the structure. The end-effector oscillates around the intended position until that energy dissipates.

That oscillation is residual vibration. And it doesn't just show up at the end of moves — the robot is lagging, overshooting, and ringing at structural frequencies throughout the entire trajectory. The controller just doesn't know it.

The real cost

We've worked with teams across truck loading, precision manufacturing, and visual inspection. The pattern is always the same: vibration quietly constrains three things that directly affect the business.

The cost of vibration and Reforge Robotics' solution to it

Throughput. Every move that ends with vibration needs settling time before the next operation can begin. In a single pick-and-place cycle, that might be 50 to 200 milliseconds. Across thousands of cycles per shift, that settling time becomes hours of lost output. For high-volume applications, we're talking about a measurable hit to unit economics.

Accuracy. If the application can't wait for the vibration to settle — because the next operation starts immediately, or the arm is in continuous motion — the robot is executing at a position offset from where the controller thinks it is. In welding, dispensing, inspection, or any path-sensitive task, this degrades quality.

Deployment time. This is the one that surprises teams the most. Vibration behavior changes with speed, payload, mounting, and even the specific robot unit. A trajectory that works cleanly on one machine may vibrate badly on another. One software manager at a robotics company told us their engineers would "perpetually spend 5% of engineering time fine-tuning control speeds for different applications." Multiply that across every deployment, and you have a serious engineering bottleneck.

Why robots vibrate

Residual vibration is not a defect. It's physics.

Every robot arm is a mechanical system with mass, stiffness, and damping at every link and joint. When the robot accelerates or decelerates, those physical properties determine how the structure responds. If the commanded motion includes acceleration changes that excite the system's natural frequencies, the structure will vibrate.

Think of it like a spring-mass-damper system — the fundamental model we use in mechanical engineering. The robot's links have mass and resist acceleration (inertia). The structure resists deformation like a spring (stiffness). And energy dissipates through joint friction and material behavior (damping). The equation of motion that governs this — mx'' + cx' + kx = F(t) — is well understood. The challenge is that every robot, in every installation, has different values for those parameters.

Three factors make this especially pronounced on modern cost-effective robots:

Lighter, more flexible structures. Affordable arms use lighter materials and less rigid construction. That lowers the natural frequencies of the system, which means even moderate speeds can excite vibration. Heavier industrial robots aren't immune — they just push the problem to higher speeds.

Lower-resolution actuators. Less expensive motors and gearboxes have more compliance, and less precise torque control. What the joint actually delivers is a noisier version of what was commanded.

Generic controller tuning. Most robot controllers ship with a single set of gains optimized for a general case. They don't account for the specific payload, mounting structure, or installation. Every deployment starts with a controller that is only approximately correct for the system it's controlling.

Why common fixes fall short

The instinctive response is to slow the robot down. Lower speeds mean lower accelerations, less energy into structural modes, less vibration. It works — but it trades throughput for smoothness. At some point, the business case for automation breaks down.

The next step is manual gain tuning. An engineer adjusts controller parameters — PID gains, acceleration limits, jerk limits — until the vibration is acceptable for a specific trajectory and payload. This works for a fixed scenario. Change the speed, change the payload, move the robot to a different mounting, and you're back to tuning.

The common thread: these approaches are either too conservative, too narrow, or too labor-intensive to scale.

Closing the gap with data from the real machine

The fundamental problem is that the controller doesn't know enough about the physical system it's controlling. It commands a trajectory based on an idealized model — or no model at all — and the real dynamics create the gap that shows up as vibration.

We solve this by calibrating the robot as installed — on its real structure, with its real payloads, using its real actuators — and changing the model with the system's configuration and payload. During calibration, sensors capture the machine's physical responses while controlled excitations expose its resonant frequencies and structural behavior. From that data, we build a precise and comprehensive dynamic model — a mathematical replica of how that specific robot actually moves, bends, and vibrates.

That model deploys directly into the control architecture. The software anticipates and neutralizes physical deviations before they happen. No manual gain tuning. No conservative speed limits. No months of trial and error.

Here's what that looks like in practice on a KUKA KR500:

The results speak for themselves: we've demonstrated larger than 85% reduction in vibration even with a 2x increase in productivity. The machine gets faster and smoother at the same time — through software alone, with zero hardware modifications.

The best part is we've built production infrastructure around it. You start with a base model that you fine-tune to your setup. Our databases store and maintain models for each robot in your fleet. And we provide enterprise-grade support as you scale.

Why this matters now

Cost-effective robots are proliferating across the general-purpose robotics market. They're cheaper, easier to deploy, and increasingly paired with AI systems that can plan, perceive, and adapt.

But intelligence doesn't eliminate motion error. A robot still has to follow the path. It still has to move quickly without shaking. It still has to repeat the same motion reliably. It still has to behave in production the way it behaved in testing.

The industry has largely solved the design and mass manufacturing of affordable robots. What it hasn't solved is making those robots perform like high-end machines. That's a software problem — and it's the one we're working on.

The robot vibrates because the controller doesn't know how the real machine moves. Fix that, and the robot stops shaking.

Reforge Robotics builds open-source motion control software that helps robots move faster, track more accurately, and behave more predictably.

Contact us to learn more about how we address residual vibration.

I recently read Poor Charlie’s Almanack by Peter D. Kaufman, which is a series of transcribed talks by the late Charles T. Munger (of Berkshire Hathaway). In these talks, Charlie lays out a peculiar process for structured thinking that he used throughout his life (in all his business and personal affairs). If you think about thinking for any amount of time, it becomes clear that most of us don’t have a structured way for general thinking. In specific domains, such as sales at a company ABC, we train employees to use certain processes when solving problems. However, when confronted with a new, complex, and interdisciplinary problem, it’s challenging to extrapolate because we don't have a latticework for general structured thinking.

Charlie sought out a structured process to think clearly. His solution was an interdisciplinary checklist that espoused all the major ideas from the major disciplines, many of which he learned on his own. The checklist includes ideas from mathematics (e.g., compound interest, Bayesian probability, inversion, etc.), science and engineering (e.g., critical mass, Darwinian evolution, backup systems, etc.), and psychology (e.g., social proof, sunk cost fallacy, etc.). When he encountered a new investment opportunity, he’d use a two-track analysis to evaluate it. First, he’d ask himself: “What are the factors that really govern the interests involved, rationally considered?” which considers the ideas that apply to the real interests, the real probabilities, and so forth. Second, he’d ask: “What are the subconscious influences where the brain is automatically making connections–which, by and large, are useful but often malfunction?” which evaluates the subconscious conclusions that people will come to due to psychological tendencies. Using his checklist, he thought about how each “big idea” might be affecting the real interests or subconscious conclusions.

This way of thinking served him well professionally and personally–to the tune of a $2.6B net worth at the time of his death. But most people don’t have a process like this. Worse, Charlie argues that most of the big ideas are really easy to grasp. Furthermore, most of them are taught in freshman (introductory) classes. Therefore, the main blocker to most people thinking in this way is that we haven’t made the effort to organize the ideas into a form that makes it easy to practice using them. Training oneself to learn all the major ideas in all the major disciplines and organize them in a latticework like Charlie’s seems like a worthwhile pursuit. We learn the best of what other people have already figured out (oftentimes through tremendous toil). Who wouldn’t want to have that? It’s like a shortcut through life and all that is required is reading, organization, and practice. What’s not to like?

The task of organizing my own latticework is one of the major intellectual pursuits of my life. This essay is my attempt to apply this thinking to Reforge Robotics. The rest of the essay follows a similar format to Talk Four in Poor Charlie’s Almanac, in which Charlie poses the hypothetical problem of starting and scaling a non-alcoholic beverage company with a $2 million investment in 1884 to be worth $2 trillion in 2034. This context allows Charlie to display his general thinking framework to answer the question of why the Coca-Cola company has been a tremendous success. I believe Charlie could have completed the same analysis to create the business plan for the Coca-Cola company in 1884 using his checklist without having the answer (i.e., the real Coca-Cola company) to analyze. In this essay, I attempt the same analysis for Reforge Robotics using the big ideas on my checklist. As with most lifelong pursuits, this analysis will be updated as I learn more. Okay, here’s the problem:

It is 2024 in Oakland. You are brought, along with 20 others like you, before a rich and eccentric Oakland citizen named Phitzer. Both you and Phitzer share two characteristics: first, you routinely use, in problem-solving, five helpful notions (shared below), and second, you know all the elementary ideas in all the basic college courses. Phitzer offers to invest $5 million, yet only take half the equity for a Phitzer Charitable Foundation, in a new corporation organized to go into the metal manufacturing business and remain in that business only, forever.

The other half of the equity will go to the person who most plausibly demonstrates that their business plan will cause Phitzer’s foundation to be worth $25 trillion 150 years later, in the money of that later time, 2174, despite paying out part of its earnings each year as a dividend. This will make the whole new corporation worth $50 trillion, even after paying out many trillions of dollars in dividends.

To get to a solution, we will use five helpful notions that Charlie uses in Talk Four.

1. It is usually best to simplify problems by deciding the big no-brainer questions first.

2. Use numerical fluency and mathematical principles to ascertain what the quantitative targets are.

3. It is not enough to think problems through forward. You must also think in reverse.

4. The best and most practical wisdom is elementary academic wisdom taken together in a multidisciplinary manner.

5. Really big effects (lollapalooza effects) come from a large combination of factors.

Here is my solution, my pitch to Phitzer, using the five notions and what every bright college sophomore should know:

"Well, Phitzer, the big no-brainer decisions that, to simplify our problem, should be made first are as follows: First, we are never going to create something worth $50 trillion by making a new machine tool that is similar to existing machine tools and fighting for market share in a brutally competitive market, so we’ll have to focus on something completely different. Today, in the metal manufacturing business, there is a large focus on selling software licenses and services. However, software and services are also not enough because many companies sell software and services for existing machine tools and none of them have achieved the level of success we desire nor demonstrate the potential to reach that level of success in the future. Therefore, we have to combine new software with a different hardware design for manufacturing machines. When combined, the software and hardware need to work together to exceed customer expectations at a significantly lower cost. This will accomplish two results that we want: 1) our customers, contract manufacturers, who are also in a very competitive market, will feel that they are missing out on a competitive advantage if they do not buy our new and different hardware and software combination; and 2) since the combined software and hardware leads to a lower cost, our competition, the existing machine tool companies, will be slow to change because selling more expensive machines results higher revenues, higher sales commissions, higher system integration fees, etc. If we quickly make and distribute our products, we will establish a large lead with our software and hardware by the time they realize that they need to copy our products. At that point, it might be too late to turnover their businesses to compete with us. It will take a long time and consistent effort to accumulate this lead, but if we are successful, the knowledge we gain will give us an advantage for some time to come.

To considerably lower costs for our customers, it’s clear that we need to make both our own hardware and software. By making our own hardware, we retain three low-hanging fruit advantages: 1) we don’t need to pay the profit margin that other companies place on their products or the taxes for each transaction; 2) we become independent and don’t need to rely on the hardware provider to give us increasing access to their software to improve our product; and 3) we can design our hardware to have the technologies necessary to maximize the usefulness of our software. Making both hardware and software has been a no-brainer strategy for some of the most profitable companies. Others, who only make software, have strong monopolies from being early entrants into the software market and setting the software standards for the hardware manufacturers. We unfortunately do not have such favorable conditions in the century-old market of manufacturing.

One argument against this integrated hardware and software approach is that building new factories to make hardware requires a lot of upfront investment. To this, I pose two rebuttals: (a) it is short-sighted to focus on the millions we save today at the expense of the billions we will save tomorrow, and (b) as we significantly reduce the cost of manufacturing machines for our customers, we can use those same machines to reduce our cost. This creates a flywheel effect: our machines help to make copies of themselves, reducing the cost of our products, which further reduces the cost of our hardware.

It’s worth mentioning that, when we start the company, we will work on the software first and use commercial hardware instead of building our own hardware immediately. This choice will enable four things: 1) we can sell the packaged software and commercial hardware to early customers that fit into our target demographic to begin learning from the market; 2) we begin earning revenue quicker and limit the burden (both in time and equity) to raise more financing; 3) we can select the best qualities of the existing hardware(s) and combine them to design our own hardware, while solving for deficiencies; and 4) we solve most of the software challenges first and understand the areas where hardware provides a better solution than software. If we build the hardware first or both at the same time, it would be impossible to achieve the aforementioned qualities. In the software-only phase of the business, we will be capital efficient by working with robot integrators and financers, to provide the integration and financing services to our customers, while we serve solely as the software provider. Although we sacrifice profits from the high-margin integration and financing businesses, this trade-off will allow us to focus more resources on designing and manufacturing our hardware prototypes and iterating quickly to a useful solution, which will be more beneficial to us in the long term.

Accordingly, the characteristics of the hardware platform to accomplish our result are that it 1) is low-cost; and 2) can be combined with software to improve performance. Industrial robot arms have these characteristics and there are many case studies that have shown their efficacy in manufacturing when general purpose robot arms are combined with advanced software. They can both increase the efficiency of overall operations by automating repetitive processes and they can be used to conduct a subset of the metal manufacturing work. Scaling these case studies broadly will be achieved iteratively through research and development. Additionally, the robot arms we design specifically for manufacturing will further improve manufacturing performance.

Now that we’ve answered the no-brainer questions, we will next use numerical fluency to ascertain what our target of $50 trillion implies. It’s difficult to make estimations about the market 150 years from now. Nevertheless, there are two reasons why the 150-year target is useful: 1) we want to work towards conditions that allow the company to long outlive us, and 2) human demand for manufactured metal products is guaranteed to be conserved and very likely to grow significantly. Today, the United States has a median individual annual income of about $40,000 per year and its people enjoy many manufactured metal goods to obtain a relatively superior quality of life compared to the rest of the world. In contrast, the median annual global income is about $3,000 and, of the 8 billion people in the world, more than 7 billion live on lower incomes than the median US income. We can guess reasonably that by 2174, the fraction of people at the equivalent of a US median income in 2024 will be greater, which guarantees sustained demand for the current level of manufacturing machine tools, likely with periods of extreme demand as one region or another experiences significant growth. Today, approximately 500,000 CNC machines are sold annually at a growth rate of 4-6%. Assuming the growth is conserved, we can expect 1.5 billion units to be shipped in 2174. Thus, if our new machine, and other imitative machines in our new market, can supply over 25% of machine tools worldwide, and, with our fanaticism about low cost, we can occupy 40% of the new market, we can sell 150 million units in 2174. Assuming a profit of only $25,000 on each unit, we can reach $3.75 trillion of free cash flow on a revenue run rate of $37.5 trillion (assuming an average unit price of $250,000). This will be enough, given our business is still growing at a good rate, to make it easily worth $50 trillion.

A big question, of course, is whether $25,000 is a reasonable profit target for 2174. And the answer is yes if we can create a product with strong universal appeal. One hundred and fifty years is a long time and the dollar will almost surely suffer monetary depreciation. Concurrently, real purchasing power of the average consumer in the world will go way up. Her proclivity to purchase manufactured goods to improve her quality of life will go up considerably faster. Meanwhile, as technology improves, the cost of our product, in units of constant purchasing power, will go down. All four factors will work together in favor of our $25,000 per machine profit target. We also expect the cost of labor to rise with the purchasing power of the average consumer and the software that comes along with our machine will help replace a portion of labor from efficiency gains. The history of software teaches us that manufacturers will be willing to pay for the increased efficiency. Therefore, even if the hardware profit is only $10,000, the rest can be made up from the price of the software. Altogether, worldwide machine tool purchasing power in dollars will probably multiply by a factor of at least 5 over 150 years. Thinking in reverse, this makes our profit-per-machine target, under 2024 conditions, a mere one-fifth of $25,000, or $5,000. This is an easy to exceed target as we start out if our new product has universal appeal.

To create a product with universal appeal, we must tackle the two intertwined challenges of large scale. First, over 150 years, we must cause a new machine tool market to assimilate about one-fourth of the world’s machine tools. Second, we must operate so that 40% of the new market is ours while our competitors are left to share the remaining 60%. These results are lollapalooza results. Accordingly, we must attack our problem by causing every favorable factor we can think of to work for us. Plainly, only a powerful combination of many factors is likely to cause the lollapalooza consequences we desire.

Let’s start by exploring the consequences of our simplifying no-brainer decision that we need to combine software with a new hardware design for machine tools. This conclusion automatically leads to an understanding of our business in proper terms. We can see from the introductory course in psychology that, in essence, we are going into the business of creating and maintaining social proof by demonstrating the benefit of the new hardware and causing a large number of conversions to our machine tool system. It is not enough for our product to perform better than legacy products. Those products benefit from social proof and the change resistance inherent in man. The only solution to create behavior we seek is a new cult-like movement with a zealous missionary following. A movement with this characteristic is the only way to convince a large group of people to do something differently who have grown accustomed to doing things a certain way. We have to employ many characteristics of other successful movements (like religions):

1. A mission that transcends the individuals/products, coupled with a missionary mandate which strongly encourages people in our movement to tell others about it.

2. Strong admiration for people who use our products.

3. Automatic word-of-mouth sharing, in social circles, about our movement. Telling a friend about our movement should be a no-brainer, universal benefit to both the teller and the hearer.

4. Special classes of community members who have some status or benefit. For example, a special class for people who exclusively use our machines and get special benefits for this exclusivity. Classes will be prominently featured so that members aspire to reach status by, for example, exclusively using our machines.

5. Physical communal aspects (e.g., a friendly feeling) by creating opportunities for our members to interact with each other and make new friends within the community, and bring their friends to meet others in the community.

To generate strong admiration for people who use our products, we will elevate their positive and moral character qualities, most of which have nothing to do with our product. We will profile qualities like their family life and their community contributions in an effort to get others to admire them greatly. We will also find already publicly admired figures and incentivize them to endorse our community. Additionally, if the qualities we focus on are associated with the larger mission of the community, the mission will be reinforced within the community. These activities will also support word-of-mouth sharing as admired persons will be talked about in circles where their values and the mission find resonance.

An important note is that the positive association must be created between admired persons and the movement/community, not our product or the company. Our product will simply benefit from an association with the movement, but it will need to demonstrate its merits to each customer independent of the movement.

Similarly, the community will have benefits (preferably free benefits) that are completely unrelated to the company or product, in addition to ones related to the company (e.g., discounts). On the company side, benefits will be integrated into a showroom as one of the physical spaces where community members interact with one another. Other benefits can include free meals, drinks, and gifts.

Since we’ve discussed forces that would favor the universal appeal of the company and product, we must now think in reverse to find forces that oppose it. What would make it more difficult to change the manufacturer’s mind? How can we ensure failure to introduce a new product and further entrench the incumbents? As usual, let’s start with the no-brainers:

1. Our technology does not live up to existing machine tool performance or is worse.

2. Our combined hardware and software system is more difficult to use than the existing manufacturing systems, creating puzzlement and stress.

3. We attempt to force new manufacturers to make a fast decision about adopting our system and we don’t give them time to think through the decision and learn about how the system works. When forced to make a fast decision, the system with social proof will be chosen over ours.

4. We focus on making comparisons to the existing systems instead of focusing on what our system can uniquely do, thereby pounding in the existing machines’ favorable qualities in user’s minds and creating avenues for debate.

5. We display our product as having weak qualities and weak associations instead of displays of strength, which are associated with rigidity and precision.

6. We highlight people who are using the existing machines and are successful and we ensure a mental association between their success and their existing machines.

7. We first focus on getting older people, who tend to hate change, to change their behavior instead of focusing on younger people who are more likely to try something new.

8. Over time, we maintain existing machine programming and workflow paradigms which makes it easy for customers to switch back to their older machines or for the old companies to create competing machines and use their reputation to recapture market share.

Avoiding factors 1-3 favors a slower rollout strategy where we validate the technology before rolling it out and give manufacturers the opportunity to gain familiarity with it. We will eliminate factors 4-6 with careful messaging to: (a) highlight only the unique features of our system and avoid making comparisons; (b) use strong names and strong qualities to describe our products; and (3) never advertise a positive story with machines that are not ours. Factor 7 is easy to eliminate by focusing on recruiting younger people first to join our movement and convincing them that our products are the future of manufacturing. Factor 8 is the most difficult to eliminate because we must lower the software adoption barrier (factor 2) for new users but quickly move existing customers to a new and improved programming and workflow paradigm. We will achieve this by maintaining multiple versions of software which change from the old to new paradigm one version at a time. The software will be compatible across versions and a process will be developed and updated to move users from the old complex software to the new simpler software over time. We may find that the new software is so simple and intuitive that our users are able to grasp it quickly, but the more likely outcome is that it will require training over time.

Well, that is my solution, Phitzer, to the problem of turning $5 million into $50 trillion even after paying out trillions of dollars in dividends. The correct strategies are clear after being related to elementary academic ideas brought into play by the helpful notions."

Most robotics startups are focused on some aspect of building better software for robots to do X, where X is some application that really needs to be automated. The prevailing notion is that all the hardware challenges in robotics have been solved. All that’s left to do is write UI software that makes the robots more performant (faster, more dexterous, etc.). Easy enough, right? Plus, now we have AI. AI will solve all the challenges that we can’t solve with classical software.

Unfortunately, this view is incorrect. The design and mass manufacturing of robots has largely been solved, but that doesn’t mean we’ve solved all the hardware challenges. The software that controls the hardware is one of the major “hardware challenges” that still needs work. Let me explain:

To mass manufacture robots for many industries, some trade-offs need to be made. One of these trade-offs is using bare-bones (e.g., PID-based) control algorithms that enable repeatable positioning but not necessarily accurate positioning, meaning that the robot will go to the same (wrong) position 99.9% of the time. Another trade-off is that the control system can’t really account for objects it will interact with in the real world. Hence, the robot’s contact with these objects can impact both its repeatability and accuracy.

When software engineers encounter these problems, you may hear them say, “robots just aren’t accurate enough” or “robots just aren't strong enough” to do X application. But that’s not the whole story; there’s just more work to be done. Clearly, the robot manufacturers can’t write custom control software for each robot to meet each user’s accuracy/object specs. They depend on users (oftentimes through system integrators) to do that themselves while they focus on the important job of churning out general-purpose robots. However, developing control software to improve the robot’s performance requires a deep understanding of robot dynamics (i.e., physics) and control theory (i.e., abstract math), which are not typical learning outcomes in software engineering coursework.

So how do we tackle this mismatch? Can it be solved or are we cursed to never have high-performance robots running great software? To start, it’s helpful to recall how we got here.

The Software Boom and Bust

Back in 2013, when I got to Stanford, the landscape of software development was quite different. The tech industry had a shortage of computer programmers and my classmates who were majoring in Computer Science (CS) were in high demand for internships, almost guaranteed to get well-paying software development engineering (SDE) roles upon graduation. This trend was a direct result of the internet boom that began in the late 90s and continued through the early 2000s, driving a massive need for skilled programmers.

During this period, internet software companies like Google, Meta (formerly Facebook), Amazon, as well as several startups were in fierce competition to attract top talent. They offered extravagant compensation packages to lure the best software engineers and started the era of six-figure salaries for entry-level engineers. The demand for these skills was insatiable, fueled by both the rapid growth of the internet and competition to prevent talent from joining other companies or starting their own.

However, today we find ourselves in a much different place. We are now a decade past that explosive growth, and the dynamics of the tech industry have shifted. There is an increasing perception that we may have too many software engineers. Driven by the demand, software engineering education (both formal and informal) grew dramatically over the past decade. For example, CS consistently ranks as the most popular engineering major at top universities across the world. The same tech companies that were on the hiring sprees last decade are cutting back, and their focus has shifted to individuals with highly specialized skills, like knowledge of advanced machine learning algorithms.

There are two key features of internet software, in particular, that contributed to the decrease in demand:

1. Scalability: Once the software was written, it could scale infinitely. The effort required to maintain and update the software was significantly less than the original work needed to build it. This scalability reduces the long-term demand for large numbers of software engineers. Additionally, the internet’s winner-take-all dynamics led to a few companies using their existing reach in one market to build bundled products that quickly grew their market share in other markets (think Google).

2. Advancements in code automation: The recent rise of large language models (LLMs), and other code automation technologies before LLMs, revolutionized the way we approach software development. These products are great at generating and maintaining code, further reducing the need for a large workforce of engineers. Automation is compounded by the fact that there’s a lot of open-source internet code to use as data to train LLMs and other tools. In contrast, there isn’t nearly as much data for other types of software (e.g., embedded systems control software).

While the demand for software engineers is certainly not disappearing, the kinds of roles available are transforming. There is a growing need for SDEs to adapt and learn how to build for different kinds of systems outside internet software and this transformation is opening new opportunities, particularly in robotics.

The Transition to Robotics

As discussed above, transitioning from developing software for digital systems to creating software for hardware systems, like robots, is challenging for traditional software engineers. This difficulty largely stems from a lack of training in the fundamental principles of physics and mechanics, which are crucial for understanding and manipulating the physical world.

However, hope is not lost. We’ve seen remarkable early examples of software engineers partnering with experts in the sciences to build innovative real-world capabilities. A prime example of this collaborative success is the research on protein folding. By combining new software algorithms (like Transformers) with decades of biological research into the structure of proteins, researchers achieved groundbreaking results (see AlphaFold from Google DeepMind and structure-informed language models from Stanford). The same synergy between different domains of expertise is also paving the way for similar advancements in robotics (see Covariant and Dexterity)! The playbook seems to be: (1) a strong understanding of the underlying nature of the problem, rooted in scientific fundamentals, then (2) the addition of elegant software to transform scientific insights into efficient code. Unfortunately, software is not good enough on its own and AI is not good enough on its own. (Heck, physics isn’t good enough on its own.)

Recently, I've noticed a growing trend of early-career SDEs expressing an interest in pivoting to robotics. This is good news. They’re becoming aware of the saturation of talent in the digital software market and are interested in the relatively untapped potential of robotics. However, at the risk of repeating myself, I would caution these engineers against the belief that software or AI alone will solve robotic automation problems. Instead, I would recommend a study of fundamental robot mechanics (you can start with Robot Dynamics and Control by Mark Spong) and finding a mechanics or controls expert to work with.

How Reforge Robotics Fits In

At Reforge Robotics, we are well-positioned to benefit from this influx of CS talent. Our team has a strong background in physics and control engineering, which complements the skills of strong software engineers to build robust robot applications.

We intend to drive advancements in robotics and automation in the manufacturing industry. Through the combination of physics-based robot control and user-centered software development, we can handle complex physical environments in manufacturing and meet the needs of our customers with software that is 10x easier to use than traditional machines.

As the value of our products for manufacturers becomes increasingly evident, we anticipate a continued surge of interest from software developers eager to build applications for manufacturing robots on our underlying architecture. We plan to build APIs for other developers to use our robot models and controllers to build software for more applications and use-cases. This model reminds me of how NVIDIA showcased the practical benefits of accelerated computing via their GPUs by enhancing computer graphics applications and subsequently built CUDA, a platform that enabled developers to write accelerated computing code. Today, many AI platforms run on NVIDIA’s chips using CUDA software. We anticipate a similar trajectory for Reforge Robotics.

Today, we are in the infancy of automation and the transition to automating physical systems presents both challenges and opportunities. The future of robotics demands a convergence of computer science and the physical sciences. This interdisciplinary approach will lead to scalable physical interactions between robots and their surroundings, particularly in the manufacturing context. By building a collaborative ecosystem where the best software engineers and physical engineers/scientists can work together, we can overcome the challenges and leverage the opportunities.

We intend to build the next generation of manufacturing systems by combining: (1) the hard-won software engineering efficiencies developed over the past decade, and (2) a modern (and historical) understanding of the physical sciences, driven by advancements in fundamental research. Reforge Robotics is committed to being a pioneer in this new era. Our strategy will not only drive advancements in manufacturing automation but also create a framework for many other industries to adopt robotic automation.