Behavior Cloning Foundations

Behavior cloningbehavior cloningBehavior cloning is supervised learning of a policy from expert demonstrations, with observations as inputs and the expert's actions as labels. The recipe predates deep learning: ALVINN (1988) used camera and laser-range inputs and trained on simulated road images before NAVLAB tests. Scaled up, it still underlies most learned manipulation policies; its known closed-loop failure mode is covariate shift. is imitation learning cast as supervised learning: a policy $\pi(a \mid s)$ is regressed onto expert actions given expert observations. In closed loop the robot's own actions choose its future inputs, breaking the i.i.d. assumption, so errors compoundcompounding errorCompounding error is the accumulation of small per-step mistakes in a sequential task. Each error moves the robot slightly off its training data, where the next is likelier, so total deviation outgrows the per-step rate. Ross, Gordon, and Bagnell bounded it for imitation learning at total cost quadratic in episode length. Action chunking reduces it by cutting sequential decisions.^{Source 7}7 Ross et al. 2010, Efficient Reductions for Imitation Learning. Stéphane Ross, J. Andrew Bagnell, AISTATS 2010 (PMLR vol. 9).

Paper figure: Alexander Khazatsky et al. / arXiv. Licence: CC BY 4.0.

Watch the error compound

Try “A teacher corrects it” to watch DAgger-style relabeling^{Source 2}2 Ross et al. 2011, A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. Stéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell, AISTATS 2011 cut the summed deviation.

Origins in driving

ALVINN (1988), a three-layer back-propagation network with camera and laser-range inputs, trained on simulated road images and was tested on NAVLAB, a modified Chevy van^{Source 1}1 Pomerleau 1988, ALVINN: An Autonomous Land Vehicle in a Neural Network. Dean A. Pomerleau, Advances in Neural Information Processing Systems 1, 1988. Later driving systems added off-centre, rotated training views^{Source 8}8 Bojarski et al. 2016, End to End Learning for Self-Driving Cars. Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski et al., 2016, or synthesized perturbations once 30 million examples proved too few^{Source 9}9 Bansal et al. 2018, ChauffeurNet: Learning to Drive by Imitating the Best and Synthesizing the Worst. Mayank Bansal, Alex Krizhevsky, Abhijit Ogale, 2018.

Photo: Firefly4342 / Wikimedia Commons. Licence: CC BY-SA 4.0.

Covariate shift, stated precisely

Training data follow the expert's state distribution $d_{\pi^*}$; the robot visits its own $d_\pi$, drifting further with each mistake^{Source 2}2 Ross et al. 2011, A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. Stéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell, AISTATS 2011. Ross and Bagnell (2010) proved a regret bound that grows quadratically in the horizon $T$^{Source 7}7 Ross et al. 2010, Efficient Reductions for Imitation Learning. Stéphane Ross, J. Andrew Bagnell, AISTATS 2010 (PMLR vol. 9). The DAgger paper restates their bound for expert-distribution error $\varepsilon$ bounding the 0-1 loss and task cost in $[0,1]$, and notes it is tight^{Source 2}2 Ross et al. 2011, A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. Stéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell, AISTATS 2011:

With log-loss, behavior cloning reaches horizon-independent sample complexity when payoff range and policy-class complexity are controlled^{Source 10}10 Foster et al. 2024, Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning. Dylan J. Foster, Adam Block, Dipendra Misra, 2024, though a tabular lower bound of order $|S|T^2/N$ survives even active expert queries^{Source 11}11 Rajaraman et al. 2020, Toward the Fundamental Limits of Imitation Learning. Nived Rajaraman, Lin F. Yang, Jiantao Jiao et al., 2020. With continuous actions, smooth deterministic imitators can suffer execution error exponentially larger in the horizon than their training error^{Source 12}12 Simchowitz et al. 2025, The Pitfalls of Imitation Learning when Actions are Continuous. Max Simchowitz, Daniel Pfrommer, Ali Jadbabaie, 2025.

Training-side problems

On demonstrations with several valid strategies, energy-based implicit policies often beat mean-squared-error and mixture-density policies^{Source 13}13 Florence et al. 2021, Implicit Behavioral Cloning. Pete Florence, Corey Lynch, Andy Zeng et al., 2021, and Diffusion Policydiffusion policyA diffusion policy is a policy generating a whole action sequence by conditional denoising diffusion instead of regressing one action. Introduced by Chi et al. in 2023, it represents multimodal demonstrations without averaging their modes. It executes only the first few steps of each sequence, then observes again and replans. learns both Push-T modes, committing to one per rollout^{Source 6}6 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023. Pan et al. (2025) attribute generative policies' benchmark gains to iterative computation instead of multimodality^{Source 14}14 Pan et al. 2025, Much Ado About Noising: Dispelling the Myths of Generative Robotic Control. Chaoyi Pan, Giri Anantharaman, Nai-Chieh Huang et al., 2025.

A single Markovian state cannot predict pauses, so single-step policies freeze, as ACT and Diffusion Policy document^{Source 5}5 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023 ^{Source 6}6 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023. Temporal context is the usual fix, yet more input information can worsen a cloned policy under shift^{Source 15}15 Haan et al. 2019, Causal Confusion in Imitation Learning. Pim de Haan, Dinesh Jayaraman, Sergey Levine, 2019, and a history-conditioned policy can cheat by copying the expert's previous action^{Source 16}16 Wen et al. 2020, Fighting Copycat Agents in Behavioral Cloning from Observation Histories. Chuan Wen, Jierui Lin, Trevor Darrell et al., 2020.

Behavior cloning at best matches its demonstrator, and robomimic found results depend on demonstration quality and checkpoint selection^{Source 17}17 Mandlekar et al. 2021, What Matters in Learning from Offline Human Demonstrations for Robot Manipulation. Ajay Mandlekar, Danfei Xu, Josiah Wong et al., CoRL 2021. Past a threshold, extra demonstrations per environment or object barely help; diversity matters far more^{Source 18}18 Lin et al. 2024, Data Scaling Laws in Imitation Learning for Robotic Manipulation. Fanqi Lin, Yingdong Hu, Pingyue Sheng et al., ICLR 2025, 2024. In Park and Levine's OCBench (October 2026), whose scripted demonstrators mimic human ones, success can keep rising as a policy overfits, and on two tasks closed-loop policies almost completely fail where action chunking works^{Source 31}31 Park et al. 2026, Behavioral Cloning Mystery. Seohong Park, Sergey Levine, 2026.

DAgger: relabel the states you actually visit

DAggerDAggerDAgger (Dataset Aggregation) is an iterative fix for behavior cloning's distribution mismatch: the expert labels states the current policy visits, and the policy retrains on the aggregate. Ross, Gordon, and Bagnell framed it as reducing imitation learning to no-regret online learning. The task-cost guarantee is existential over the learned sequence and assumes learning and expert-mixing conditions. Linear excess-cost scaling needs a horizon-independent recovery-cost factor u, which can be O(T) at worst. (Dataset Aggregation) labels the states the learner visits, reducing imitation learningimitation learningImitation learning is learning a policy from expert demonstrations instead of a reward signal, by supervised fitting of recorded state-action pairs. Pomerleau's ALVINN (1988) trained on simulated road images before steering NAVLAB; the recipe still underlies most learned manipulation. Its structural weakness: training states come from the expert while deployment visits the learner's own, the mismatch DAgger was designed to repair. to no-regret online learning^{Source 2}2 Ross et al. 2011, A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. Stéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell, AISTATS 2011.

Infinite-sample Theorem 3.2 assumes a bounded, strongly convex surrogate loss and expert mixing $\beta_i \leq (1-\alpha)^{i-1}$, with $\alpha$ independent of $T$. After $N=\tilde{O}(uT)$ iterations some learned policy satisfies the bound below, where $u$ bounds the rise in expert cost-to-go after one action^{Source 2}2 Ross et al. 2011, A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. Stéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell, AISTATS 2011:

Excess cost is linear in $T$ only if $u$ is independent of $T$; in the worst case $u$ is $O(T)$. The guarantee is existential; Algorithm 3.1 returns the policy best on validation. Finite-sample Theorem 3.4 holds with probability at least ${1-\delta}$ for empirical loss $\hat{\varepsilon}_N$, $m=O(1)$ trajectories per iteration and $N=O(u^2T^2\log(1/\delta))$^{Source 2}2 Ross et al. 2011, A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. Stéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell, AISTATS 2011.

Off-policy DART injects noise into demonstrations so the supervisor shows recoveries, gaining 62% on average over behavior cloning at grasping in clutter^{Source 19}19 Laskey et al. 2017, DART: Noise Injection for Robust Imitation Learning. Michael Laskey, Jonathan Lee, Roy Fox et al., 2017, and Diffusion Meets DAgger synthesizes out-of-distribution samples, reaching 80% pushing success from 8 demonstrations against behavior cloning's 20%^{Source 20}20 Zhang et al. 2024, Diffusion Meets DAgger: Supercharging Eye-in-hand Imitation Learning. Xiaoyu Zhang, Matthew Chang, Pranav Kumar et al., 2024.

Labeling arbitrary failure states is unnatural for teleoperators. In HG-DAgger the human takes over when needed, steers back to a safe, stable region and hands back control, labeling the recovery^{Source 3}3 Kelly et al. 2019, HG-DAgger: Interactive Imitation Learning with Human Experts. Michael Kelly, Chelsea Sidrane, Katherine Driggs-Campbell et al., arXiv v2 (11 March 2019). BC-Z gathered such interventions by shared autonomy and did 24 unseen tasks at 44% average success^{Source 21}21 Jang et al. 2022, BC-Z: Zero-Shot Task Generalization with Robotic Imitation Learning. Eric Jang, Alex Irpan, Mohi Khansari et al., 2022, and Compliant Residual DAgger takes delta corrections without pausing the policy, raising base success by 64% on four tasks^{Source 22}22 Xu et al. 2025, Compliant Residual DAgger: Improving Real-World Contact-Rich Manipulation with Human Corrections. Xiaomeng Xu, Yifan Hou, Chendong Xin et al., 2025.

Physical Intelligence's Recap coaches a vision-language-action policy with corrections at visited states^{Source 4}4 Amin et al. 2025, π*0.6: a VLA that Learns from Experience. Ali Amin, Raichelle Aniceto, Ashwin Balakrishna et al., 2025. RaC found imitation plateaus far below perfect despite thousands of demonstrations, and its recovery-and-correction data beat the prior best with 10× less collection^{Source 23}23 Hu et al. 2025, RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction. Zheyuan Hu, Robyn Wu, Naveen Enock et al., 2025. In September 2026, Abhishek Gupta, Yi Shao and colleagues used DAgger to distill a state-based RL teacher into an RGB student; their RebarSim seated rebar at 1.4 mm clearance in 137 of 150 real rollouts, and in simulation pure behavior cloning in place of the DAgger mixture cut success from 96.1% to 68.9%^{Source 29}29 Sun et al. 2026, Visual Sim-to-Real Learning for Robotic Insertion under Geometric Variations: Application to Rebar Installation. Tao Sun, Beining Han, Patrick Yin et al., 2026.

NVIDIA's Recova (October 2026) has an agent develop recovery skills in a digital twin of each task. On four real tasks at 20 trials each, DAgger fine-tuning raised mean success from 23.8% to 77.5%, and recovery skills raised it to 87.5%^{Source 30}30 Liu et al. 2026, Recova: Agent-Guided Failure Recovery for Autonomous Robotic Manipulation. Isabella Liu, An-Chieh Cheng, Johan Bjorck et al., 2026.

Beyond DAgger

Action chunking and exploratory data collection each provably avoid exponential compounding through control-theoretic stability^{Source 24}24 Zhang et al. 2025, Action Chunking and Exploratory Data Collection Yield Exponential Improvements in Behavior Cloning for Continuous Control. Thomas T. Zhang, Daniel Pfrommer, Chaoyi Pan et al., 2025. A generative imitator, such as a diffusion policy, over a stabilizing controller can match the per-step distribution of expert trajectories^{Source 25}25 Block et al. 2023, Provable Guarantees for Generative Behavior Cloning: Bridging Low-Level Stability and High-Level Behavior. Adam Block, Ali Jadbabaie, Daniel Pfrommer et al., 2023.

GAIL's adversarial imitation needs no expert queries^{Source 26}26 Ho et al. 2016, Generative Adversarial Imitation Learning. Jonathan Ho, Stefano Ermon, 2016, nor does DARP's nearest-neighbour retrieval, 15 to 46% better than behavior cloning^{Source 27}27 Pfeifer et al. 2026, Difference-Aware Retrieval Policies for Imitation Learning. Quinn Pfeifer, Ethan Pronovost, Paarth Shah et al., 2026. HIL-SERL adds real-world RL, reaching near-perfect success in 1 to 2.5 hours^{Source 28}28 Luo et al. 2024, Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning. Jianlan Luo, Charles Xu, Jeffrey Wu et al., 2024.