Behavioral Cloning

Behavioral cloning is the approach to imitation learning that reduces control to a supervised learning problem. Given a dataset of states or observations paired with the actions an expert took in them, a model is trained to predict the expert action from the observation, and the trained predictor is then deployed as the control policy. Nothing about environment dynamics or rewards enters the training objective. The method appears in the literature as "behavioral cloning", "behavior cloning" and the original British "behavioural cloning", and the abbreviation BC is standard in robotics.

The attraction is that behavioral cloning inherits the whole toolkit of supervised learning: stable optimization, off-the-shelf architectures, no simulator, no reward function to design, and no environment interaction during training. Collect demonstrations, fit a regressor or classifier, deploy. That simplicity is why the large robot policies released between 2022 and 2026, among them RT-2, Octo, OpenVLA and the Physical Intelligence π0 family, are trained primarily by behavioral cloning on demonstration data.

The weakness is equally structural, and it is not a matter of insufficient data or a weak model class. A supervised learner is trained on the state distribution the expert visits, but at deployment it visits the state distribution its own actions induce. Those two distributions diverge as soon as the policy makes its first mistake, and the divergence feeds on itself. Stephane Ross, Geoffrey Gordon and J. Andrew Bagnell put the consequence sharply: a classifier that errs with probability epsilon under the expert's state distribution can make as many as T squared times epsilon mistakes over a T-step episode under its own distribution, because "as soon as the learner makes a mistake, it may encounter completely different observations than those under expert demonstration, leading to a compounding of errors" ^[1].

The supervised reduction

Behavioral cloning assumes access to a dataset of trajectories generated by an expert. Each trajectory is a sequence of observations and the actions the expert chose. Training minimizes a surrogate loss between the model's predicted action and the expert's recorded action: mean squared error for continuous action spaces such as joint velocities or steering angles, cross-entropy for discretized or tokenized actions, and more recently a denoising or flow-matching objective when the policy models a distribution over action sequences rather than a point estimate.

Three design choices distinguish practical implementations. The first is what the model conditions on: a single frame, a short history of frames, proprioceptive state, or a natural-language instruction. The second is what it outputs: a single next action, or a chunk of many future actions predicted at once. The third is how the output distribution is represented, which matters because human demonstrations are multimodal. Two operators shown the same scene will often pick different valid actions, and a model trained with a squared-error loss averages them into an action neither demonstrator would take. Diffusion Policy was motivated in part by this problem, representing the visuomotor policy as a conditional denoising diffusion process in order to handle multimodal action distributions gracefully, and reported an average 46.9% improvement over prior methods across 12 tasks from 4 manipulation benchmarks ^[2].

Origins

The usual first example is ALVINN, the Autonomous Land Vehicle In a Neural Network built by Dean Pomerleau at Carnegie Mellon. The system described in the 1988 NIPS proceedings was a three-layer backpropagation network with 1,217 input units (a 30x32 video retina, an 8x32 laser range finder retina, and one road-intensity feedback unit), 29 hidden units and 46 output units, 45 of which encoded turn curvature. It was trained on 1,200 simulated road snapshots for 40 epochs and drove the CMU NAVLAB, a modified Chevy van, at half a meter per second along a 400-meter path through a wooded part of campus ^[3].

That first version is not strictly behavioral cloning: the training images came from a road generator, not a human driver. The genuinely relevant paper is Pomerleau's 1991 follow-up in Neural Computation, which introduced "training on-the-fly", teaching the network to imitate a human driver under real driving conditions using the current camera image as input and the direction the person was steering as the target. Pomerleau identified the distribution-shift problem in that paper, roughly two decades before it was formalized: "since the human driver normally steers the vehicle down the road center during training, the network will never be presented with situations where it must recover from errors" ^[4].

His fix anticipated later augmentation practice. Each captured image was laterally shifted in software, seven times left and seven times right in 0.25-meter increments, creating 14 additional images in which the vehicle appears displaced from the road center and 15 training exemplars in all, with the steering label adjusted for each shift. A buffer of 200 recent road scenes was maintained, with 15 replaced per cycle (the 10 lowest-error exemplars plus 5 chosen at random). About 50 cycles were needed, roughly five minutes on a Sun-4 while a person drove the test road at about 4 mph. The trained network ran at 25 images per second and drove at up to the NAVLAB's maximum of 20 mph ^[4].

The name came from a separate lineage in symbolic machine learning. Michael Bain and Claude Sammut's paper "A Framework for Behavioural Cloning" in the Machine Intelligence 15 volume describes the method as one where "the logged data from skilled, human operators are input to an induction program which outputs a control strategy for a complex control task", producing situation-action rules. Bain and Sammut also recorded the failure mode in plain terms: because the clone has no representation of goals, "when a situation occurs which is outside of the range of experience represented in the training data, the clone can fail entirely" ^[5].

The best-known experiment in that tradition is "Learning to fly" by Claude Sammut, Scott Hurst, Dana Kedzier and Donald Michie, presented at the Ninth International Conference on Machine Learning in 1992. They modified a flight simulator to log the control actions a human subject took while flying a fixed flight plan, then ran an induction program over the log and installed the resulting decision tree as an autopilot. Three pilots flew 30 flights each, producing a dataset of roughly 90,000 recorded events ^[5]. The pipeline is exactly the modern one, with a decision tree in place of a neural network.

Compounding error and distribution shift

The formal account arrived with Ross and Bagnell's 2010 AISTATS paper "Efficient Reductions for Imitation Learning", which showed that the naive supervised approach yields "a regret bound that grows quadratically in the time horizon of the task" because the learned policy influences the future test inputs on which it is evaluated ^[6]. This is distribution shift of a self-inflicted kind: the training and test input distributions differ not because the world changed but because the learner acts.

The following year Ross, Gordon and Bagnell introduced DAgger (Dataset Aggregation), which turns imitation into no-regret online learning ^[1]. The algorithm is short. Initialize an empty dataset D and an arbitrary policy. At each of N iterations, roll out a mixture policy that plays the expert with probability beta_i and the current learned policy otherwise; collect the visited states; label every one of them with the action the expert would have taken; aggregate those labels into D; retrain on all of D. Return the best iterate on a validation set. Because the states come from the learner's own induced distribution but the labels come from the expert, the aggregated dataset progressively covers exactly the situations the policy will actually encounter. Where plain behavioral cloning gives cost on the order of T squared times epsilon, DAgger gives cost linear in T ^[1].

The empirical results were blunt. On steering a kart in Super Tux Kart, the supervised baseline did not improve with more data "because most of the training laps are all very similar and do not help the learner to learn how to recover from mistakes it makes", while DAgger produced a policy that never fell off the track after 15 iterations ^[1]. DAgger and its variants remain the reference answer to compounding error, but they carry a cost that has limited adoption in robotics: they require an interactive expert who can be queried at arbitrary states, which is expensive and, for a human teleoperator watching a robot in an unfamiliar configuration, awkward to provide.

Later theory has complicated the simple picture. Nived Rajaraman and co-authors proved a minimax lower bound of order |S| H squared / N for imitation learning from N expert trajectories in episodic MDPs, matched up to a log factor by an algorithm with no dependence on the number of actions, which establishes that the quadratic horizon dependence is not merely an artifact of a loose analysis ^[7]. In the other direction, Dylan Foster, Adam Block and Dipendra Misra argued in 2024 that horizon-independent sample complexity is achievable in offline imitation learning when the range of cumulative payoffs is controlled and the policy class has bounded supervised learning complexity, and that behavior cloning with logarithmic loss can match the horizon dependence of online methods under dense rewards. Their conclusion is that the gap between offline and online imitation learning is smaller than the classic quadratic bound suggests ^[8].

Other failure modes

Compounding error is not the only problem. Pim de Haan, Dinesh Jayaraman and Sergey Levine documented causal confusion: because behavioral cloning fits a discriminative model rather than a causal one, it can latch onto observed variables that correlate with the expert's action without causing it. The counterintuitive result is that "access to more information can yield worse performance", since a richer observation gives the model more spurious correlates to exploit, and those correlates break under distribution shift ^[9].

Two further ceilings are inherent to the formulation. A cloned policy is bounded by its demonstrator: with no reward signal, there is no mechanism by which it can become better than the behavior it was fit to. And demonstration data is expensive in a way that web text is not, because every trajectory costs a person real time at the controls. In the ALOHA experiments each episode took a human operator 8 to 14 seconds depending on task complexity, which at 50 Hz control works out to 400-700 recorded timesteps per demonstration ^[10].

Behavioral cloning in modern robot learning

Two things changed after 2022. Sequence models made it practical to predict chunks of future actions rather than one step at a time, and cheap teleoperation rigs made demonstration collection tractable for academic labs.

The ALOHA system, published by Tony Zhao, Vikash Kumar, Sergey Levine and Chelsea Finn in 2023, is the clearest example of both. It is a bimanual setup costing under 20,000 USD, where the operator backdrives two small WidowX leader arms whose joints are synchronized with two larger ViperX follower arms, with teleoperation and recording at 50 Hz ^[10]. The accompanying policy, Action Chunking with Transformers (ACT), trains a conditional variational autoencoder over action sequences. The paper is explicit about why: predicting a chunk of k actions and executing it as one unit "reduces the effective horizon of the task by k-fold, mitigating compounding errors" ^[10]. The ablation is direct evidence that horizon length drives behavioral cloning failure. Averaged over four settings, success rose from 1% with no chunking (k = 1) to 44% at k = 100, then tapered slightly toward open-loop control at k = 200 and k = 400 ^[10].

With 50 demonstrations per task, about 10 minutes of data, ACT reached 88% on sliding open a Ziploc bag and 96% on slotting a battery, while the four prior imitation learning baselines made no progress past the first stage of either task. On the harder tasks, ACT reached 84% on opening a translucent condiment cup, 92% on putting on a shoe, 64% on preparing tape, and only 20% on threading a velcro cable tie, the one task for which 100 demonstrations rather than 50 were collected ^[10]. The ALOHA hardware, together with its Mobile ALOHA successor, became a common reference platform for demonstration collection, and the surrounding tooling was consolidated in Hugging Face's LeRobot library, which provides models, datasets of human-collected demonstrations and simulated environments with a stated focus on imitation learning and reinforcement learning approaches shown to transfer to real hardware ^[11].

Scale followed. RT-1 was trained on more than 130,000 episodes gathered over 17 months with 13 robots covering over 700 tasks, and ran at 3 Hz; it succeeded on 97% of seen instructions, 76% of never-before-seen instructions, 83% of distractor-robustness tasks and 59% of background-robustness tasks ^[12]. RT-2 replaced the bespoke architecture with a vision-language model co-fine-tuned on robot trajectories, expressing actions as text tokens so that actions and language share a format, and evaluated the result over 6,000 trials ^[13]. The Open X-Embodiment collaboration then pooled data from 22 robot embodiments across 21 institutions covering 527 skills ^[14].

Scroll sideways for more →

Octo is a transformer policy trained on 800,000 trajectories from Open X-Embodiment, instructable by language command or goal image, fine-tunable to robot setups with new sensory inputs and action spaces within a few hours on standard consumer GPUs, and evaluated across 9 robotic platforms ^[15]. OpenVLA is a 7-billion-parameter vision-language-action model built on Llama 2 with a fused DINOv2 and SigLIP visual encoder, trained on 970,000 real-world robot demonstrations; it outperformed the 55-billion-parameter RT-2-X by 16.5 percentage points of absolute task success across 29 tasks with seven times fewer parameters, and beat Diffusion Policy by 20.4 points in fine-tuning settings, with model checkpoints, fine-tuning notebooks and the training codebase released publicly ^[16]. π0 from Physical Intelligence put a flow-matching action expert on top of a pretrained vision-language model and demonstrated laundry folding, table cleaning and box assembly ^[17].

A July 2025 study from the Toyota Research Institute pushed back on how loosely such systems are usually evaluated, arguing that meaningful measurement of real-world performance had become the limiting factor in the field. The team extended the Diffusion Policy paradigm into what it calls Large Behavior Models and assessed them through blind, randomized trials. Multitask pretraining made the policies more successful and more robust, let them learn new tasks from a fraction of the single-task data, and produced performance that rose predictably with pretraining scale and diversity ^[18].

End-to-end driving

The ALVINN idea returned as autonomous driving research scaled. NVIDIA's 2016 DAVE-2 system trained a convolutional neural network to map raw pixels from a single front-facing camera directly to steering commands, running at 30 frames per second on a DRIVE PX computer, and learned to handle local roads with and without lane markings, highways, parking lots and unpaved roads without explicit lane-detection or path-planning modules ^[19]. The paper's framing, that the network "automatically learns internal representations of the necessary processing steps such as detecting useful road features", is the end-to-end argument in one sentence ^[19].

The modern version substitutes a multimodal foundation model for the CNN. Waymo's EMMA, first posted in October 2024 and revised through September 2025, is built on a Gemini foundation and maps raw camera data directly to planner trajectories, perception objects and road graph elements, representing all of its non-sensor inputs and all of its outputs as natural language text so that a single model handles multiple tasks through task-specific prompts. It reported state-of-the-art motion planning on nuScenes and competitive results on the Waymo Open Motion Dataset ^[20]. The underlying learning signal is unchanged since 1991: a human drove, the model predicts what the human did.

Contrast with inverse RL and reinforcement learning

Behavioral cloning, inverse reinforcement learning and reinforcement learning differ in what they infer from the data.

Scroll sideways for more →

Inverse reinforcement learning, formulated by Andrew Ng and Stuart Russell at ICML 2000, attacks the problem of extracting a reward function from observed optimal behavior rather than copying the behavior itself ^[21]. Pieter Abbeel and Ng's apprenticeship learning built on it, treating the expert as maximizing a reward expressible as a linear combination of known features and showing that the learned policy attains performance close to the expert under the expert's own unknown reward ^[22]. The argument for the detour is generalization: a reward function is a more compact and more transferable description of a task than a policy, and it can be re-optimized under new dynamics. The argument against it is cost, stated plainly in the Generative Adversarial Imitation Learning paper, which observed that recovering the cost function and then extracting a policy from it "is indirect and can be slow", and proposed instead to learn the policy directly through a generative adversarial objective ^[23].

In practice the field has converged on behavioral cloning for pretraining and reinforcement learning for the last mile. Physical Intelligence's π*0.6, released in November 2025, is the clearest statement of that division. Its method, Recap (RL with Experience & Corrections via Advantage-conditioned Policies), runs three stages: demonstrations, then coaching in which expert teleoperators intervene with real-time corrections, then autonomous reinforcement learning using a value function that scores actions by advantage. Training on autonomous experience "more than doubles the throughput on some of the hardest tasks" and "can decrease failure rates by 2x or more" on making espresso drinks, folding laundry and assembling boxes ^[24].

Developments in 2025 and 2026

The clearest trend is co-training on data that is not robot demonstrations. π0.5, released in April 2025, combines multi-robot data, high-level semantic prediction and web data into hybrid multimodal examples that mix image observations, language commands, object detections, semantic subtask prediction and low-level actions; the headline result was long-horizon manipulation such as cleaning a kitchen or bedroom "in entirely new homes" ^[25]. NVIDIA's GR00T N1, released in March 2025, uses a dual-system design pairing a vision-language module with a diffusion transformer for motor actions, jointly trained end to end on real robot trajectories, human demonstration videos and synthetic data ^[26].

Gemini Robotics 1.5 and Gemini Robotics-ER 1.5, announced by Google DeepMind on 25 September 2025, added an explicit reasoning step before acting, generating an internal sequence of natural-language analysis, and demonstrated cross-embodiment transfer, moving behaviors trained on ALOHA 2 hardware to Apptronik's Apollo humanoid and to Franka bi-arm robots without embodiment-specific specialization ^[27].

At the small end, SmolVLA (June 2025), from Hugging Face, was trained on community-collected data from affordable robot platforms rather than academic or industrial corpora, designed to train on a single GPU and run on consumer GPUs or even CPUs while remaining competitive with vision-language-action models ten times larger ^[28]. At the frontier, Physical Intelligence's π0.7 (April 2026) reported early signs of compositional generalization: it folded laundry on a bimanual UR5e system for which no laundry-folding data had been collected, at a success rate the company says matches the zero-shot success rate of its own expert teleoperators attempting the same transfer ^[29]. Distilling experience generated during Recap training back into the general model let a single π0.7 match or exceed the success rates and throughput of the task-specific π*0.6 specialists on laundry folding, espresso making and box building. Its training mixture included human data and autonomous episodes collected by running earlier policies, alongside conventional demonstrations ^[29].

The common thread across all of these is that behavioral cloning is treated as the pretraining stage rather than the whole method. Demonstrations establish a competent prior; chunked action prediction shortens the effective horizon; heterogeneous co-training broadens coverage; and reinforcement learning or on-robot correction closes the gap that compounding error opens.

See also

• Imitation learning
• Distribution shift
• Diffusion policy
• Vision-language-action model
• Reinforcement learning
• Robot learning

References

1. ^Ross, S., Gordon, G. J., and Bagnell, J. A. "A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning." AISTATS 2011. proceedings.mlr.press/...ross11a.pdf (preprint: arxiv.org/...1011.0686)
2. ^Chi, C., Xu, Z., Feng, S., Cousineau, E., Du, Y., Burchfiel, B., Tedrake, R., and Song, S. "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion." arXiv:2303.04137, 7 March 2023. arxiv.org/...2303.04137
3. ^Pomerleau, D. A. "ALVINN: An Autonomous Land Vehicle in a Neural Network." Advances in Neural Information Processing Systems 1, NIPS 1988. proceedings.neurips.cc/...d43bbf5bbe87fb-Paper.pdf
4. ^Pomerleau, D. A. "Efficient Training of Artificial Neural Networks for Autonomous Navigation." Neural Computation 3(1), pp. 88-97, 1991. publications.ri.cmu.edu/...pomerleau_dean_1991_1.pdf
5. ^Bain, M., and Sammut, C. "A Framework for Behavioural Cloning." Machine Intelligence 15. cgi.cse.unsw.edu.au/...MI15.pdf
6. ^Ross, S., and Bagnell, D. "Efficient Reductions for Imitation Learning." AISTATS 2010, PMLR 9, pp. 661-668. proceedings.mlr.press/...ross10a
7. ^Rajaraman, N., Yang, L. F., Jiao, J., and Ramachandran, K. "Toward the Fundamental Limits of Imitation Learning." arXiv:2009.05990, 13 September 2020. arxiv.org/...2009.05990
8. ^Foster, D. J., Block, A., and Misra, D. "Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning." arXiv:2407.15007, 20 July 2024. arxiv.org/...2407.15007
9. ^de Haan, P., Jayaraman, D., and Levine, S. "Causal Confusion in Imitation Learning." arXiv:1905.11979, 28 May 2019. arxiv.org/...1905.11979
10. ^Zhao, T. Z., Kumar, V., Levine, S., and Finn, C. "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware." arXiv:2304.13705, 23 April 2023. arxiv.org/...2304.13705 (project page: tonyzhaozh.github.io/aloha)
11. ^Hugging Face. "LeRobot" documentation. huggingface.co/...index
12. ^Brohan, A., et al. "RT-1: Robotics Transformer for Real-World Control at Scale." arXiv:2212.06817, 13 December 2022. arxiv.org/...2212.06817 (project page: robotics-transformer1.github.io)
13. ^Brohan, A., et al. "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control." arXiv:2307.15818, 28 July 2023. arxiv.org/...2307.15818
14. ^Open X-Embodiment Collaboration. "Open X-Embodiment: Robotic Learning Datasets and RT-X Models." arXiv:2310.08864, 13 October 2023. arxiv.org/...2310.08864
15. ^Octo Model Team, Ghosh, D., Walke, H., Pertsch, K., et al. "Octo: An Open-Source Generalist Robot Policy." arXiv:2405.12213, 20 May 2024. arxiv.org/...2405.12213
16. ^Kim, M. J., Pertsch, K., Karamcheti, S., et al. "OpenVLA: An Open-Source Vision-Language-Action Model." arXiv:2406.09246, 13 June 2024. arxiv.org/...2406.09246
17. ^Black, K., Brown, N., Driess, D., et al. "π0: A Vision-Language-Action Flow Model for General Robot Control." arXiv:2410.24164, 31 October 2024. arxiv.org/...2410.24164
18. ^TRI LBM Team. "A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation." arXiv:2507.05331, 7 July 2025. arxiv.org/...2507.05331
19. ^Bojarski, M., Del Testa, D., Dworakowski, D., et al. "End to End Learning for Self-Driving Cars." arXiv:1604.07316, 25 April 2016. arxiv.org/...1604.07316
20. ^Hwang, J.-J., Xu, R., Lin, H., et al. "EMMA: End-to-End Multimodal Model for Autonomous Driving." arXiv:2410.23262, 30 October 2024. arxiv.org/...2410.23262
21. ^Ng, A. Y., and Russell, S. "Algorithms for Inverse Reinforcement Learning." ICML 2000. ai.stanford.edu/...icml00-irl.pdf
22. ^Abbeel, P., and Ng, A. Y. "Apprenticeship Learning via Inverse Reinforcement Learning." ICML 2004. ai.stanford.edu/...icml04-apprentice.pdf
23. ^Ho, J., and Ermon, S. "Generative Adversarial Imitation Learning." arXiv:1606.03476, 10 June 2016. arxiv.org/...1606.03476
24. ^Physical Intelligence. "π*0.6: a VLA that Learns from Experience." 17 November 2025. pi.website/...pistar06
25. ^Physical Intelligence. "π0.5: a Vision-Language-Action Model with Open-World Generalization." arXiv:2504.16054, 22 April 2025. arxiv.org/...2504.16054
26. ^NVIDIA. "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots." arXiv:2503.14734, 18 March 2025. arxiv.org/...2503.14734
27. ^Google DeepMind. "Gemini Robotics 1.5 brings AI agents into the physical world." 25 September 2025. deepmind.google/...-agents-into-the-physical-world
28. ^Shukor, M., Aubakirova, D., Capuano, F., et al. "SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics." arXiv:2506.01844, 2 June 2025. arxiv.org/...2506.01844
29. ^Physical Intelligence. "π0.7: a Steerable Model with Emergent Capabilities." 16 April 2026. pi.website/...pi07