Imitation Learning Explained, With the Astribot S1
The Astribot S1 is trained by imitation learning: a person performs the task, the robot records what happened, and a neural network learns to reproduce it. That differs from the process most people picture, in which a programmer specifies waypoints and joint angles by hand. This article is about the method rather than the machine. For the hardware, see our full write-up of the Astribot S1 humanoid robot.
What imitation learning means on a working robot
An imitation-learned robot runs a policy: a function taking whatever the robot can sense, usually camera images plus its own joint positions, and outputting the next commanded action. Training it is supervised learning. You collect paired observations and actions from a person doing the task correctly, then fit a model that predicts the action a human would have taken.
That framing separates it from the alternatives. Classical motion planning computes a trajectory from an explicit geometric model of the world. Reinforcement learning discovers behaviour by trial and error against a reward function, which suits locomotion and suits nothing you cannot afford to fail at ten thousand times. Imitation learning needs no reward engineering and no world model, but somebody has to show the robot what good looks like. Astribot states that training the S1 on a task typically takes 10 to 50 demonstrations: small for one narrow task, large across every variation a household contains.
Why hand-coded motion planning stalls on manipulation
Traditional industrial robots are programmed, not trained, and for good reason. In a fixture-heavy cell the part sits in a known pose, the gripper is purpose-built, and the same motion repeats for a million cycles. A programmed trajectory is faster, cheaper and more auditable than any learned policy under those conditions.
Manipulation outside a fixture breaks all of those assumptions. The object arrives in an arbitrary pose. It may be deformable, so it has no single rigid transform to plan against. Success depends on contact forces and slip conditions that resist analytical modelling. Writing a controller for "fold this towel" means writing a model of cloth.
Precision hardware does not rescue the software. The S1 is specified at approximately 0.1 mm position repeatability with force control to roughly 0.1 N resolution. Repeatability tells you the arm lands where the controller asked. It says nothing about whether the controller knew where to ask, and that is what imitation learning supplies.
How a demonstration gets captured
Three methods dominate. Kinesthetic teaching means backdriving the arm through the motion by hand: cheap and intuitive, but limited to lightweight arms, and the operator's hands occlude the cameras meant to record the scene. Teleoperation means driving the robot through a controller while its own sensors record the result. Astribot uses VR teleoperation for the S1, and this is the workhorse method, because recorded actions already sit in the robot's action space. Video of humans is cheapest to gather and hardest to use, since a human arm has different kinematics.
Exoskeleton rigs occupy a useful middle ground. The Elephant Robotics myController S570 exoskeleton controller is a wearable dual-arm rig with 12 degrees of freedom, six per arm, and 570 mm of reach per arm. It reads joint angles from 4096-count magnetic encoders, streams them at up to 100 Hz, and adds roughly 30 to 50 ms of latency. It lists at $1,299 at the time of writing.
Sample rate and latency decide how good the dataset is. Sample too slowly and fast contact events fall between frames, so the policy never sees the moment that mattered. Add too much latency and the operator overcorrects, baking their compensation for the delay into the data as though it were part of the task.
From recordings to a usable dataset
Raw demonstrations are not a dataset. Camera streams have to be time-aligned with joint states, because images that lead or lag by a few frames teach a subtly wrong relationship. Camera positions must stay fixed across sessions, or the policy learns the camera rather than the task. Failed attempts have to be discarded or labelled, since behaviour cloning treats the whole training set as correct.
Teams wanting a bench for the perception half of this, without tying up a research humanoid, often use a fixed station. The Elephant Robotics Portable AI Educational WorkStation packs compute, input devices and a working surface into a 600 by 400 by 250 mm case weighing 11 kg without an arm fitted, and ships with 30 lesson plans and five visual recognition algorithms.
What behaviour cloning learns, and where it breaks
Behaviour cloning is the simplest form of imitation learning and still the most widely deployed. Its central weakness has a name: compounding error, sometimes called covariate shift. Training data contains only states a competent person passed through. At run time the policy makes a small error, landing slightly outside the training distribution. The next prediction is worse, pushing it further out, and the deviation grows with every step. The robot fails because it has wandered somewhere nobody showed it.
Mitigations all widen that distribution: collect corrective demonstrations from the states the policy actually drifts into, perturb the robot mid-demonstration so recoveries appear in the data, predict a short sequence of future actions per inference step instead of one, and vary lighting, object pose and clutter.
- Long-horizon tasks. A ten-step task is not ten times harder than a one-step task. Error accumulates, and one early mistake invalidates everything after it.
- Distribution shift. A new countertop height, different lighting, an unfamiliar mug. Trivial for a person, frequently fatal for a policy trained in one kitchen.
- Recovery behaviour. If the demonstrator never dropped the cup, the policy has no idea what to do about a dropped cup.
- Precision beyond the demonstrator. A machine repeatable to 0.1 mm will not exceed the accuracy of the demonstrations it was given.
- Unobserved state. Demonstrators use tactile feedback the capture rig may never record, and what is absent from the observation stream cannot be learned.
- Per-task data cost. Ten to 50 demonstrations per task is manageable. The same across several hundred tasks is an operations problem, and currently the main brake on general-purpose deployment.
The Astribot S1 as a worked example
Read the S1's specification with the training method in mind and the design choices line up. It carries two 7-DoF arms on a 4-DoF articulated torso, mounted on a 3-DoF omnidirectional wheeled base, standing about 1.7 m tall. Each arm handles up to 3 kg at full extension. End-effector speed peaks at 10 m/s, the base moves at about 1.5 km/h, and a charge supports 4 to 6 hours of continuous operation with a 2 to 3 hour recharge. It sells for $100,000.
The wheeled base is the most instructive choice. A legged humanoid learns balance and manipulation at once, and every disturbance from the arms feeds back into the locomotion controller. Rolling removes that coupling, reducing the learning problem to manipulation alone. The 4-DoF torso buys back most of the workspace legs would have reached, so a policy can finish a task without the base repositioning. Fewer moving parts means fewer demonstrations to cover them.
Comparable research humanoids running the same workflow
Labs that specifically want the locomotion coupling back choose a legged platform. The LimX Dynamics Oli EDU humanoid robot stands 165 cm, weighs up to 55 kg with its battery, and carries 31 active degrees of freedom: seven per arm, six per leg, three at the waist and two in the neck. Single-arm payload is 3 kg and runtime is about two hours per charge. It lists at $72,000.
For imitation-learning work the important item on its spec sheet is not a joint count. It is the complete URDF supplied for sim-to-real transfer, with support for NVIDIA Isaac Sim, MuJoCo and Gazebo. An accurate model lets you pretrain on synthetic demonstrations and fine-tune on a far smaller set of real ones. That is the standard answer to the data-cost problem.
| Platform | Price | Mobility | Single-arm payload | Runtime per charge |
|---|---|---|---|---|
| Astribot S1 Humanoid Robot | $100,000 | 3-DoF omnidirectional wheeled base | Up to 3 kg | 4 to 6 hours |
| LimX Dynamics Oli EDU Humanoid Robot | $72,000 | Two legs, 6 DoF each | 3 kg | About 2 hours |
Frequently asked questions
What is imitation learning in robotics?
Imitation learning trains a robot from recorded examples of a person performing a task, rather than from hand-written instructions. The recordings pair what the robot could sense with the action taken at that moment, and a network is fitted to predict the second from the first. No reward function and no explicit physical model of the objects are required.
How many demonstrations does a robot need to learn a task?
That depends on how much variation the task contains. Astribot states the S1 typically needs 10 to 50 demonstrations per task. A tightly constrained task with a fixed object in a fixed pose sits at the low end. Varied objects, lighting or starting positions need considerably more, because a policy only handles conditions it has seen.
Is teleoperation the same as imitation learning?
No, though the two are usually used together. Teleoperation is a person directly driving the robot in real time. Imitation learning is training a model on the recordings teleoperation produces. A teleoperated robot stops working the moment the operator leaves. An imitation-trained robot runs the learned policy on its own, which is the point of collecting data.
Why does imitation learning beat hand-coded motion planning for manipulation?
Hand-coded planning needs an explicit geometric and physical model of whatever is being manipulated. That is achievable for rigid parts in fixtures and impractical for cloth, food, cables or cluttered household objects. Imitation learning skips the model. It learns the mapping from what the cameras see to what the arms should do, which is often easier to obtain than the physics.
What is the biggest weakness of behaviour cloning?
Compounding error. Training data contains only states a skilled person visited, so once the policy makes a small mistake it lands in a state it was never trained on, where its next prediction is worse still. The deviation grows step by step. Fixing it means collecting demonstrations of recovery from the exact states the policy drifts into.
Choosing a platform for demonstration-driven work
Pick the platform that matches the problem you are studying. If the research question is manipulation, a wheeled base removes a variable you would otherwise fund with data. If it is whole-body control, you need the legs and the simulation work that comes with them. Confirm before purchase that the manufacturer ships a usable URDF, a documented SDK and a teleoperation path. Browse the humanoid robots available at RoboticsSelect, including the Astribot S1 dual-arm humanoid. As an authorized dealer we ship every unit with its full manufacturer warranty.