Behavior cloning treats robot control as an ordinary supervised learning problem: collect a dataset of (observation, action) pairs from a demonstrator, then train a model to predict the action given the observation, the same way you would train an image classifier. No reward function, no environment model, no exploration — the entire learning problem is regression or classification, using tools that are decades old and well understood. That simplicity is exactly why it is the first approach most manipulation teams reach for.
Why it is the default for manipulation
Reinforcement learning needs a reward signal and, usually, thousands of trials to converge, which is expensive and often unsafe on real hardware. Behavior cloning needs neither: it trains on however many demonstrations you can collect, using the same supervised learning infrastructure — batches, gradient descent, standard architectures — that the rest of machine learning already runs on. For a new manipulation task, a team can go from a handful of teleoperated demonstrations to a trained policy in an afternoon, which is not true of any RL approach on physical hardware. That is why behavior cloning is the baseline every serious manipulation stack starts from, even when it later adds other techniques on top.
Distribution shift and compounding error
The problem behavior cloning inherits from being ordinary supervised learning is that supervised learning assumes train and test data are drawn from the same distribution — and a deployed control policy breaks that assumption by construction. Every action the policy takes changes the next observation it sees, so once the policy makes an error, it moves into a state slightly outside its training distribution, where its predictions are less reliable, and the process repeats. Ross and Bagnell's analysis of this problem (Efficient Reductions for Imitation Learning, AISTATS 2010, and A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning, AISTATS 2011, with Geoffrey Gordon) showed that naive behavior cloning's error grows quadratically with task horizon under this feedback loop — a mechanism covered in more depth under imitation learning. Practically, this is why a behavior-cloned policy that looks fine for the first few seconds of a rollout can drift into failure by the end.
The multimodality problem: why naive MSE fails
A second, distinct failure shows up even with distribution shift ignored. Human demonstrations of the same task are rarely a single consistent motion — a demonstrator might approach a grasp from the left in one episode and the right in the next, both correct. A policy trained with mean squared error regression on this data learns to predict the average of the two options, which can be a motion that satisfies neither and lands the end effector somewhere between two valid grasps rather than at either one. This is a structural property of the loss function, not a lack of data: more demonstrations of the same two-mode behavior only sharpen the wrong average.
Three families of fix have emerged, all still fundamentally behavior cloning, differing only in how they represent the output distribution:
- Discretized action tokens. Bin the continuous action space and predict a categorical distribution over bins, so the model can represent distinct modes instead of averaging them — the approach RT-1 and RT-2 use, encoding actions as tokens from a fixed vocabulary.
- Diffusion policies. Model the action distribution as a denoising diffusion process conditioned on the observation, which naturally represents multimodal distributions rather than collapsing them to a mean. Chi et al.'s Diffusion Policy: Visuomotor Policy Learning via Action Diffusion (RSS 2023) demonstrated this directly against MSE and other baselines.
- Action chunking. ACT (Action Chunking with Transformers, Zhao et al., RSS 2023, the policy behind the ALOHA bimanual rig) predicts a short sequence of future actions at once using a conditional variational autoencoder rather than one action per step, which both represents multimodality within the chunk and shortens the effective planning horizon, reducing how much compounding error accumulates before the policy replans.
An earlier line of work, Implicit Behavioral Cloning (Florence et al., CoRL 2021), reframed the policy as an energy-based model that scores action candidates rather than regressing to a point estimate — one of the first demonstrations that the choice of output representation, not the demonstrations themselves, was the source of the multimodality problem.
What data properties actually make it work
Behavior cloning is only as good as three properties of its dataset, independent of which output representation is used: consistency (a demonstrator executing the task with a stable, learnable strategy rather than improvising a different approach each time), coverage (the states a near-competent policy will actually encounter, including recoveries from small errors, not only clean nominal executions), and sufficient repetition per variation for the model to separate a real pattern from one demonstrator's idiosyncratic noise. None of these is solved by collecting more of the same; they are properties a data collection protocol has to deliberately design for.
In practice this means a collection session is not just "run the task N times." It means holding demonstrator technique consistent enough that the policy can learn a single strategy, while deliberately varying the scene — object pose, lighting, background clutter — so the strategy is learned as a function of what's observed rather than memorized against a fixed setup. It also means treating a near-miss and its recovery as more valuable than another clean repetition, since the recovery is the only place in the dataset that teaches the policy what to do once it has already drifted slightly off course — which is precisely the state distribution compounding error pushes a deployed policy into.