BLOG

October 28, 2026

Apollo's Journey in Human-Centric Robotics and Whole Body Control

Apollo was built to take on the demanding, repetitive, and hazardous work that wears people down, and whole body control is what makes it possible.

Mastery of Whole Body Control is Foundational to a Robot for Humans

Apollo was created with a simple purpose: to help people. By supporting human teams with physically demanding, repetitive, or potentially hazardous work, Apollo can help reduce strain, improve safety, and give people more time to focus on higher-value tasks. Delivering on each of these responsibilities requires mastery of whole body motion - whether moving between locations to do work, reaching for objects on the ground or the back of a shelf, or efficiently and fluidly repositioning to do a task more consistently. This is done via whole body control, which is the software layer that defines the physical API for the system. It handles the particular physics of any given embodiment - whether wheeled or legged - and …

Years of experience working with our partners have highlighted that simply being able to demonstrate a task is not enough - the way in which Apollo does the task matters. Modern AI approaches for loco-manipulation rely on the ability of the underlying system to be able to move smoothly and with a consistency that minimizes distribution shift between data collection and evaluation across an entire fleet. It is also critical to reduce unnecessary motion to maximize energy efficiency and throughput, but also to establish trust in cooperative environments alongside humans by ensuring the robot motion is intuitive and predictable for those humans sharing the same space.

These insights drove us to overhaul our controls stack in favor of motion imitation-based whole body control, trained on human data via reinforcement learning in massively parallel simulation and transferred zero-shot to hardware. This blog post covers that journey and some of our key observations along the way in “building the engine” for a human-first robot. It is not meant to market an “AI breakthrough” but rather simply provide a transparent technical update on our progress and how we think about robotics, whole body control, and what it takes to create Robots for Humans.

‍

‍

Human Motion is the Building Block for Humanoid Motion

At a high level, whole body control breaks down into two objectives - execute some task (e.g. track a base velocity, put the hands in a specific pose, follow a whole-body reference), while also achieving some other desired properties such as maintaining balance, minimizing actuator efforts, or moving with a particular style. These desired properties can sometimes be encoded with simple hand-crafted heuristics, but this manual reward engineering can be difficult to scale to all the motions Apollo needs to do - particularly for style which can be difficult to define with a few simple equations.

An alternative way to encode these properties is with motion imitation of human data. The core concept here is that humans have developed an effective way to achieve the same tasks - particularly for the tasks we care about like moving through and interacting with environments already optimized for humans - and therefore these heuristics can be replaced with data-driven approaches which incentivize similarity to human motion. Thankfully, human motion data is relatively easy to obtain, whether from motion capture, video, animation, or existing third party datasets. Apptronik has been investing each of these modalities to maximize not just the data quantity but also coverage over the behaviors we care about.

‍

‍

Critical to this process is the ability to retarget the human motion onto the Apollo embodiment with exceptional quality. This process leverages the decades of model-based control experience of our team - the same trajectory optimization techniques which formed the bedrock for our previous online model-predictive controllers now power our retargeting pipelines. This allows us to ensure that the motion data the robot is emulating is artifact-free and dynamically feasible for any of Apollo’s embodiments, which raises the ceiling of the types of motions we can perform. This is critical for customer use cases with strict throughput requirements which require mastery of the physics of robot motion.

‍

Driving Performance by Optimizing Iteration Speed

A well curated motion dataset creates a solid foundation for training RL policies, but translating these motions into stable controllers that can do all the tasks we care about at the level of quality we want requires a great training recipe. Our philosophy is that such a training recipe is unlikely to be “secret sauce”, as new state-of-the-art techniques and implementations are constantly emerging from the field at an expotential rate due to AI-driven development. Therefore our approach to policy training is not to optimize for a particular recipe, but rather optimize for what we call the “robotics iteration loop”:

  • Generate a new idea - either from prior observations or new breaking results
  • Implement the idea
  • Train a new policy
  • Test in sim
  • Test in hardware
  • Merge the code
  • Repeat at each level

We’ve found that the speed at which we can exercise this loop directly correlates to the performance of the system, and thus have deliberately accelerated each element of this loop. This means investing in integrated AI workflows for rapid code implementation, on-prem and cloud compute for model training coupled with our own in-house job management software to efficiently schedule training runs on available compute and track their progress, and integrated data visualization tools to quickly assess motion quality and identify gaps in model performance. We’ve even optimized how we schedule and coordinate robot time to ensure ideas can be tested as quickly as possible. Together, these investments mean we can take an idea from either a new paper or a morning brainstorm and have it running on a robot within the same day. That speed of performance improvement ensures we’re able to continuously improve the capabilities that Apollo supports.

‍

Our Holistic Approach to Hardware Transfer

These tools allow us to quickly train performance policies in simulation, but translating results from simulation to hardware zero shot - known as “sim-to-real” - requires getting all the little details right. Our team has a saying, “hardware is where great ideas go to die.” It is easy to have a great idea that works well in theory or even simulation, but misses some detail about real world hardware execution that breaks core assumptions - control loops can introduce jitter in unintuitive ways, mechanisms can complicate how limits are modeled, or a mis-calibrated sensor can quietly give incorrect data.

Our approach to sim-to-real is thus not to try to match the simulator to the robot, but rather to design the whole system so the two agree. That means leverging our expertise as a full-stack robotics company to design robots that are both capable and easy to model, validated for great sim-to-real throughout the design and bringup process. It also means rigorously testing our software deployment stack to ensure it runs code on robot hardware exactly the same way it does in sim. Doing this at scale requires good overall system health and sim to real performance at individual robot and fleet levels - we’ve created extensive tooling to both characterize and improve system health over time to ensure consistency of each platform.

This expertise and full-stack ownership is particularly valuable at the actuator level. It is hardly overexaggeration to say a robot is mostly just a collection of actuators, and our experience over a decade of actuator design have taught us critical lessons in how to design actuators well for good hardware transfer. We optimize motor and sensor selection, transmission strategy, firmware implementation, and validation procedures all with the goal of making the actuators capable, predictable, and simulatable. We’ve refined this approach such that we’re typically able to transfer simulation results to new hardware platforms first try, zero shot.

‍

Putting the Pieces Together

We’ve built on these foundational elements to improve the capabilities of Apollo 2 in two key areas - steerable natural locomotion and customer-focused whole body data collection. Our natural locomotion improvements leverage newly-created human locomotion datasets and algorithmic advances to yield a steerable walking behavior that demonstrates reliable and predictable motion across a range of velocities. This is primarily useful when Apollo must walk across a site to do work in a new location, or to return to a base station to swap a battery and continue working.

‍

‍

Our whole body control improvements also enable a new approach to continual improvement on customer use cases. Because we can collect human motion data of our customer use cases and quickly add it to our training pipeline, we can continuously improve the performance of the underlying control system over time on behaviors our customers care about. Since this physical API defines both the quality of data collected via on-robot methods such as teleoperation or RL as well as the quality of model performance trained on that data, this capability ensures that physical AI deployed on Apollo is able to scale with our customer needs.

‍

‍

This vision is how we plan to enable performant, reliable, and human-centric physical AI, but its just part of a broader journey in creating Robots for Humans. We’re looking for others who share that vision to join the journey with us.