Overview

Imitation learning teaches an agent to act by using examples of how a person or expert system behaves, often when specifying a reward is difficult. Behavior cloning directly maps observations to actions, while DAgger adds expert corrections on states the learner encounters, and inverse or adversarial imitation learning infers what behavior to match. Current methods also use diffusion policies, retrieval, world models, and visual, language, or tactile data for tasks such as robot manipulation, autonomous driving, and multi-agent coordination, with benchmarks including LIBERO, RoboMimic, and CARLA. A central challenge is that small errors can lead agents into states missing from demonstrations; improving recovery and safety while learning efficiently from limited, noisy, or conflicting examples remains an open tension.

Latest papers

287 papers37 papers in the last four weeks, up 6% on the four weeks before. 0.3% of all new papers.

Nov 13, 2026

Imitation Bootstrapped Multi-Agent Reinforcement Learning

Demonstrations are an effective way to accelerate single-agent reinforcement learning (RL), but extending this idea to multi-agent reinforcement learning (MARL) is non-trivial as it is difficult and often impossible for a single person to teleoperate multiple agents simultaneously. Round-Robin Behavior Cloning (R2BC) addresses this by letting a single demonstrator control one agent at a time while the others execute their current policies. However, the resulting imitation policies are limited by the quality of the demonstrations. We propose Imitation Bootstrapped Multi-Agent Reinforcement Learning (IBMARL), a framework that extends Imitation Bootstrapped Reinforcement Learning (IBRL) to cooperative teams and uses R2BC to train the imitation policies. At each time step, every agent receives one action proposal from its imitation policy and one from its RL policy, and an ensemble of centralized critics selects among the per-agent combinations of the two. A gated imitation penalty, a soft Q-filtered behavior cloning loss, regularizes each RL policy toward its imitation policy only where the critics value the imitation action more, so that after training each agent executes its RL policy alone, without the centralized critics. On four cooperative multi-agent RL benchmarks, IBMARL surpasses its imitation policies 2.62.6-4.2×4.2\times faster than demonstration-based baselines on two tasks, is the only method to surpass it on the other two, and exceeds the demonstrator itself on two of the four. With real human demonstrations IBMARL still surpasses imitation learning alone, and does so more than twice as fast as the best baseline.
Nov 13, 2026

Imitation Policies Learn a Proprioceptive Anchor: A Predictive Law for Perception Noise and Its Closed Loop Cost

Imitation policies for robot manipulation are often trained on noisy object positions, and they learn to discount the noisy reading. We show how much they discount it, and why. Because demonstrators steer toward the object, the robot's own state carries information about where the object is. A trained policy combines this proprioceptive anchor with the reading and weights the reading by α=V/(V+s2)α= V/(V + s^2), where s2s^2 is the reading noise and VV the anchor's error variance. VV belongs to the learner and is far smaller than a linear model of the demonstrations implies, which is why classical errors-in-variables (EIV) regression overestimates αα. Measured once, VV predicted ten held-out sensitivities across two robomimic tasks, two learners and four new noise levels, with mean absolute error 0.025 against 0.37 for classical EIV, and a parameter-free bias test rules out the simpler explanation that the policies just overestimate the noise. The discount does not carry over to deployment, where the policy produces its own state: closed-loop sensitivity to a persistent reading offset exceeded αα in all 15 policies that passed the registered gates, although the stricter registered prediction, that all would follow the offset fully, held for 8 of them. Post hoc, lower αα went with lower success, steeply for diffusion, and a registered prospective test confirmed this for diffusion on Can at two new noise levels: the measured αα predicted success before any rollout, within 0.03 (95% intervals ±0.08\pm 0.08), although training noise predicts it equally well. Every test was registered before its data, including the classical benchmark that failed, and all other analyses are labelled post hoc; all results are in simulation with synthetic Gaussian noise.
Nov 12, 2026

Inner Product Regularization for Imitation Learning

Robots that learn from demonstrations should extrapolate their policy despite state perturbations such as sensor noise, scene variation, and compounding errors. One approach to this robustness is regularization. Prior regularizers consider each state on its own and limit how sensitive the policy is to any perturbation at that state. By contrast, we explore how the policy's output shifts when given the same perturbation at different states. Our insight is that, locally, the robot should respond to state perturbations in a consistent manner: e.g., if changing an irrelevant object position does not impact the policy at one state, that same change should not shift the policy output at a nearby state. Applying this insight, we develop inner product regularization (InneR). We show that the inner product on action space induces an inner product on state space through the policy Jacobian (the Fisher information for stochastic policies), and that the null space of this inner product governs which state perturbations have a first-order effect on the policy. By clustering states whose inner products are aligned, averaging the inner products within each cluster, and minimizing the smallest eigenvalues of that average, InneR shapes the learned policy so that nearby states share a common null space. InneR enforces consistent responses to perturbations without limiting policy expressiveness, and can be appended to existing imitation learning pipelines for deterministic or stochastic policies. In simulated experiments, InneR makes learned policies more robust to irrelevant state dimensions and shifts between training and test distributions, whereas norm penalties often perform worse than behavior cloning. On a real robot arm performing three manipulation tasks, InneR raises the average success rate of a diffusion policy from 18% to 57% without additional data.
Nov 12, 2026

An Integrated Architecture for Data-Efficient Real-Time Neural Imitation of Nonlinear Time-Optimal Model Predictive Control

Nonlinear time-optimal model predictive control (TOMPC) enables highly performant control of constrained systems, but its computational cost prohibits real-time use on fast dynamical systems. Neural network imitation of the TOMPC control law can reduce online computational costs, yet existing work mostly targets quadratic cost formulations in simulation and addresses the challenges of neural controllers in isolation. This work imitates a long-horizon nonconvex TOMPC with hard terminal constraints, whose high computational cost limits training-data generation. We present (i) self-informed sampling (SIFS), a data-efficient expert-sampling strategy that uses a learned time-to-go to query the expert only where the policy deviates from time-optimal behavior, (ii) a complete training and control architecture that jointly addresses dataset correlation, suboptimal expert labels, covariate shift, constraint satisfaction and model mismatch, and (iii) real-world validation on a Furuta pendulum. SIFS matches standard dataset aggregation performance in roughly one-third of the training time and with less than half the training data. The resulting controller matches the region of attraction of the expert TOMPC while raising the control frequency from 1.4Hz to 1kHz. On hardware, the controller swings up and stabilizes the pendulum along trajectories closely matching the expert TOMPC solution.
Nov 11, 2026

Imitating the Imperfect: Offline-to-Online Robust Imitation Learning from Heterogeneous Demonstrators

Imitation learning (IL) typically relies on large-scale offline demonstrations collected from multiple human or algorithmic demonstrators. However, most existing approaches assume these demonstrators are homogeneous or near-optimal experts, a convenient but unrealistic assumption in real-world applications. In practice, demonstrations are often collected from heterogeneous demonstrators with unknown and varying levels of expertise, resulting in highly mixed-quality data. In this work, we study a challenging yet practical setting of imitation learning from such heterogeneous imperfect demonstrators. We propose Latent Expertise and Optimality Scoring based Imitation Learning (LEOS-IL), a principled framework that jointly learns (i) a state-action optimality scoring model and (ii) latent expertise levels for each demonstrator from unlabeled suboptimal offline demonstrations. The learned scoring model is then integrated into an online policy optimization procedure, where the agent is trained to maximize estimated optimality scores while its online rollouts are iteratively leveraged to refine the scoring model, thereby mitigating distributional shift during online adaptation. We further provide theoretical guarantees for optimality score identification and conditional optimal occupancy and policy recovery under explicit sufficient conditions. Experiments on continuous control benchmarks demonstrate that our approach consistently outperforms existing baselines across highly heterogeneous and low-quality demonstration regimes.
Nov 9, 2026

Making Strides in Symbolic Imitation Learning: Offline Demonstrations, Online Tuning

Deep reinforcement learning (DRL) methods have been the dominant paradigm in continuous control tasks due to their efficacy in function approximation, while interpretable evolutionary methods have lagged largely behind for the opposite reason. In this work, we introduce Symbolic Imitation Learning, a framework for distilling an oracle policy through an offline dataset into a decision list with linear model motion primitives. We show that this, in conjunction, with online Augmented-Random Search (ARS) tuning of the constants and motion primitives approximates the performance of the oracle across 20 benchmark RL tasks and exceeds the oracle in 11 of those 20 tasks while simultaneously producing parsimonious, interpretable policies that domain experts can better understand.
Nov 9, 2026

RAIL: Representation-Aligned Imitation Learning for Student-Compatible Teacher Policies

Reinforcement learning (RL) from raw sensory inputs such as images or onboard sensors can be challenging due to high-dimensional observations and sparse rewards. A common strategy is to train a teacher policy with access to privileged state information and then distill it into a student policy that acts from raw inputs alone. However, when the teacher relies on information unavailable to the student, exact imitation may be impossible, creating an irreducible imitation gap. Existing approaches typically mitigate this mismatch through careful reward shaping or additional reinforcement learning on the student policy after the teacher has been trained. Instead, we introduce Representation-Aligned Imitation Learning (RAIL), which addresses the mismatch during teacher training itself. RAIL learns a latent representation shared across teacher and student observations using contrastive learning, and trains the teacher policy directly in this space. Because the teacher policy is constrained to operate on this shared space, it learns behaviors that are reproducible from student observations. This substantially reduces the imitation gap while preserving task performance. Across multiple environments, RAIL outperforms strong baselines without reward modification or post-hoc student fine-tuning, and surpasses direct reinforcement learning from raw observations. The learned representations further enable zero-shot transfer to new tasks through teacher-only training. Project site: https://merajmammadov.com/rail/.
Nov 7, 2026

Explicitly capturing demonstrator multi-modality for distillation into diffusion policies

The diversity of a distilled diffusion policy's outputs matters beyond how it impacts the policy's success rate. In particular, downstream applications like test-time search and policy improvement require exploration. Thus, we want to distill multi-modal demonstrators into diffusion policies which reproduce their conditional action distributions. Standard data collection for behavior cloning applies a sliding window to demonstration episodes, which pairs unique observation histories with actions. These datasets can be fit by a policy that is either multi-modal, or unimodal with high Lipschitz constant to capture a diversity of actions seen at nearby observation samples. Prior work finds that diffusion policies trained on sparse data learn the latter, memorizing one action per observation. We adapt vine sampling to behavior cloning data collection to resolve this ambiguity: at some states the demonstrator visits, query it kk times from exactly the same observation. On a synthetic task, we study how the number of actions per observation, kk, affects the learned policy, and find that larger kk yields a policy that better matches the demonstrator's conditional distribution, including its multi-modality and diversity. On a 2D truck unloading task, vine sampling likewise leads to a policy closer to the demonstrator and substantially improves both best-of-NN and expert iteration performance.
Nov 7, 2026

Online Imitation Learning for Stabilizing Vlasov--Poisson Plasmas

We study the problem of controlling instability in the Vlasov--Poisson plasma systems, a fundamental challenge for nuclear fusion control. Motivated by the partial observability of plasma states in practice, we investigate imitation learning algorithms that learn from a fully-observable expert controller, but operate under macroscopic measurements. We theoretically establish a separation between online and offline imitation learning --- even a small behavior cloning error leads to error compounding that exponentially amplifies instability over time. By way of contrast, the online imitation learning loss can polynomially control the stability of rollout trajectories. We then propose a DAgger-style algorithm that learns the stabilizing controller under partial observability. Simulation results on a 1D Vlasov--Poisson system demonstrate that our algorithm can effectively stabilize plasma instabilities over longer time horizons than behavior cloning.
Nov 7, 2026

MoE-DAgger: Transition-Aware Router-Entropy Gating for Selective Human Correction in Phase-Structured Manipulation

Selective data acquisition in Dataset Aggregation (DAgger) on robots requires deciding which learner rollouts to correct and when to intervene. In manipulation with ordered phases, router entropy from a phase-supervised soft mixture-of-experts (MoE) is insufficient: the same value can reflect either a normal adjacent-phase handoff or support for nonadjacent or earlier phases. We propose MoE-DAgger, which uses causal router history to distinguish these cases without providing the policy with a phase label, progress variable, or hard phase command during execution. The acquisition rule allows an adjacent transition to settle, but requests correction when high entropy persists outside such a transition, the transition times out, or routing persistently favors an earlier phase. Under matched active human-control budgets, MoE-DAgger achieves 84.0% task success on Green and 78.0% on Red, exceeding phase-unaware filtered entropy by 22 percentage points on Green and 26 points on Red, with no observed collisions in 50 rollouts per task. On 50 shared intervention-free traces per task, 83.3% and 80.0% of its first proposals fall on failed rollouts within predefined failure windows when 15 and 25 traces are selected, compared with 70.0% and 72.0% for temporal action-chunk disagreement and 63.3% and 60.0% for ensemble disagreement. Shadow replay and an active no-veto ablation indicate that suppressing queries at normal adjacent handoffs is a key contributor to the improved allocation.
Nov 7, 2026

Task-Critical Part Alignment Improves Imitation of Object Interactions

This paper presents enhancements to the DINOBot framework for imitation learning of object-object interactions. We introduce two key contributions aimed at improving the accuracy and efficiency of alignment in imitation-based robotic tasks. First, we integrate RANdom SAmple Consensus (RANSAC) for outlier rejection during the alignment phase, effectively mitigating the impact of noisy depth data and spurious feature matches. This improves robustness when transferring demonstrations to real-world applications. Second, we refine the semantic correspondence process by focusing on task-relevant interaction regions, ensuring that the learned representation prioritizes critical object features that influence successful manipulation. We evaluated our approach in a range of real-world manipulation tasks, including picking, insertion, shaking, and pouring, using a robotic system. The experimental results demonstrate an overall improvement of 13.8% in task success rates compared to the baseline, along with faster convergence in the alignment process. These results highlight the effectiveness of our enhancements in improving precision and efficiency for object-object interaction imitation in robotics.
Nov 7, 2026

Model-Based Imitation Learning For Monocular Vision-Based Navigation

Reliable navigation with a single monocular limited field-of-view camera is challenging because the lack of depth measurements and limited spatial awareness complicate collision checking. Existing approaches to tackle these challenges fall into two categories. Model-based planners can incorporate explicit collision-avoidance guidance which leads to better generalization but are often computationally expensive and require manual tuning of the algorithmic parameters. In contrast, end-to-end approaches trained by imitation learning can be flexible but typically encode safety only implicitly and thus can suffer from high collision rates in novel scenarios. In this paper, we present a middle ground through model-based imitation learning. The model-based part is a learned collision model that can map a sequence of past images and a candidate control sequence to a distribution of worst-case obstacle clearance. This predictive distribution in turn is used to define a collision-risk cost in a trajectory optimization formulation. The imitation component uses human demonstrations to learn scene-dependent balancing of collision risk with other task costs such as goal progress. We also use it to learn the algorithmic parameters (e.g., a warm start) of our optimizer to accelerate the solution process. We show that our approach can demonstrate highly reliable navigation in cluttered environments on robots with different embodiments. We also demonstrate a substantial reduction in collision rate over recent end-to-end and model-based baselines.
Nov 6, 2026

DexDICE: Reconciling Safety and Performance via Dual-Expert Offline Imitation Learning

Ensuring high returns while providing strong safety guarantees in offline imitation learning (IL) is a fundamental challenge: existing approaches either rely on explicitly specified cost functions, which are difficult to design for complex real-world safety constraints, or require expert demonstrations that are both safe and high-performing, which are often too costly to collect. We introduce dual-expert imitation learning, which leverages two complementary data sources: (i) safety-compliant demonstrations that faithfully follow safety guidelines but may be suboptimal, and (ii) performance-oriented demonstrations that achieve high returns but may violate safety guidelines. To address this problem, we propose DexDICE (Dual-expert offline imitation learning via stationary DIstribution Correction Estimation). Instead of conventional distribution constraints, DexDICE combines DICE-style offline learning with safe support constraints, enabling agents to exploit high-return behaviors while remaining within safety-compliant regions. We prove that the ideal solution satisfying this constraint exactly has zero unsafe-region visitation with high probability. We evaluate DexDICE on a re-curated DSRL benchmark and validate it on a real-world mobile-robot navigation task trained directly from human teleoperation demonstrations, where it substantially outperforms baselines and remains robust to data scarcity and quality degradation.
Nov 6, 2026

Mitigating Covariate Shift in Behavioral Cloning via Robust Stationary Distribution Correction

We consider offline imitation learning (IL), which aims to train an agent to imitate from the dataset of expert demonstrations without online interaction with the environment. Behavioral Cloning (BC) has been a simple yet effective approach to offline IL, but it is also well-known to be vulnerable to the covariate shift resulting from the mismatch between the state distributions induced by the learned policy and the expert policy. Moreover, as often occurs in practice, when expert datasets are collected from an arbitrary state distribution instead of a stationary one, these shifts become more pronounced, potentially leading to substantial failures in existing IL methods. Specifically, we focus on covariate shift resulting from arbitrary state data distributions, such as biased data collection or incomplete trajectories, rather than shifts induced by changes in dynamics or noisy expert actions. In this paper, to mitigate the effect of the covariate shifts in BC, we propose DrilDICE, which utilizes a distributionally robust BC objective by employing a stationary distribution correction ratio estimation (DICE) to derive a feasible solution. We evaluate the effectiveness of our method through an extensive set of experiments covering diverse covariate shift scenarios. The results demonstrate the efficacy of the proposed approach in improving the robustness against the shifts, outperforming existing offline IL methods in such scenarios.
Nov 6, 2026

SUGARv2: Learning Diverse Visual Humanoid Loco-Manipulation Skills from Synthesized Interaction References

Scaling up visual humanoid loco-manipulation skill acquisition requires reducing the human effort needed as tasks and spatial configurations expand, particularly in demonstration collection and task-specific training design. We present SUGARv2, a unified framework for learning diverse visual humanoid loco-manipulation skills through interaction reference synthesis and a shared training pipeline. To obtain diverse references at low cost, we adapt reusable task templates to different spatial configurations, synthesizing coordinated robot and object motions. To reduce task-specific reward design and training effort, we use reference-based imitation as a common learning objective across tasks. We first train privileged teachers to execute these references, then train visual policies through teacher distillation and reinforcement learning with the same reference-based mimic rewards, using diverse object instances and visual domain randomization to support visual generalization and sim-to-real transfer. Across five humanoid loco-manipulation tasks, SUGARv2 achieves a mean success rate of 91.0% in simulation. Deployment on a real humanoid robot further demonstrates successful sim-to-real transfer for whole-body manipulation. These results demonstrate the framework's potential for learning diverse visual humanoid loco-manipulation skills.
Nov 4, 2026

Which Physical Tactile Cues Matter for Contact-Rich Imitation Learning? A Controlled Real-Robot Study

Vision-based tactile sensors capture multiple aspects of physical interaction within a single image stream, including contact appearance, geometry, indentation depth, and tangential deformation. However, existing visuo-tactile policies often directly process tactile observations as images, making it difficult to determine which physical cues contribute to policy learning. This paper presents a controlled real-robot study that compares physically interpretable tactile signals using the LeRobot platform, enhanced with our visuo-tactile perception system. From the same tactile stream, we derive tactile RGB images, normal/depth maps, and marker-flow fields, representing contact appearance, geometry, and tangential deformation, respectively. Using identical task demonstrations and the same policy backbone, we evaluate the individual and combined contributions of these tactile cues. We further compare tactile encoders and visuo-tactile fusion strategies. Evaluation across 720 real-robot trials yields four key findings: (1) tactile perception substantially improves the overall success rate from 57.8% to 78.9%, although its contribution varies across task stages; (2) marker flow is the most effective individual cue, achieving a success rate of 74.4% and playing a particularly important role in regulating established contact; (3) cue-factorized encoding consistently improves performance, highlighting that different physical cues require tailored processing techniques; and (4) although the effectiveness of tactile fusion is task-dependent, cross-attention offers a lightweight solution with performance comparable to that of pretrained tactile encoders.
Nov 3, 2026

GUILD: Guided Iterative Data Collection for Imitation Learning

Imitation learning policies in robotics are trained on datasets collected through costly real-world teleoperation. Often, the data-collection process is unguided, relying on the teleoperator's intuition, resulting in data that contributes little to overall policy performance. Recent work has shown that curating such unguided datasets can substantially improve policy performance, but these approaches operate by filtering existing datasets after collection and therefore do not prevent the acquisition of uninformative data. In this work, we propose GUILD, a novel Guided Data Collection (GDC) method that actively reasons about which data should be collected to maximize downstream policy performance. We introduce an iterative framework that alternates between 1) task proposal and data collection and 2) data evaluation to close the loop. For task proposal, we leverage semantic priors from vision-language models to identify relevant environments, object configurations, and goals for data collection. However, since semantic relevance does not imply policy improvement, we perform data evaluation. By leveraging predictive data attribution we approximate how the collected trajectories influence policy performance on a target deployment setting. This signal is then used as feedback to guide subsequent data collection, closing the data collection loop. We evaluate our approach in simulation and real robot settings, demonstrating that guided data collection substantially improves sample efficiency and deployment performance compared to passive data collection strategies, multiple baselines and experienced roboticists. More information at https://robin-lab.cs.utexas.edu/guild2026/
Nov 3, 2026

VARP: A Visual Adaptive Rearrangement Planner for Non-Monotone Object Rearrangement Planning

Object rearrangement planning in cluttered environments is often non-monotone, requiring objects to be temporarily relocated before reaching their final destinations. Although search-based planners can solve such tasks, they incur substantial computational cost and require expensive replanning when the environment changes. We propose the Visual Adaptive Rearrangement Planner (VARP), which learns non-monotone rearrangement directly from visual state representations generated from expert demonstrations. VARP predicts pick-and-place actions from the current and goal visual state representations, and incorporates an Action Refinement Strategy (ARS) to improve action feasibility and a Progress Enforcement Strategy (PES) to mitigate progress stagnation. By continuously updating actions from the latest visual state representations, VARP adapts online to disturbances without replanning from scratch or performing online combinatorial search. In challenging rearrangement benchmarks, VARP achieves a 98.3% success rate with only 0.4 s planning time per task, 97.5% success under execution-time disturbances, and 94.5% under continuous severe disturbances, while generalizing effectively to unseen object counts. These results demonstrate that VARP provides a scalable and robust solution for non-monotone object rearrangement.
Nov 2, 2026

KlaskTron-IL: A Contact-Rich Adversarial Imitation Learning Benchmark with a Validated Sim-to-Real Bridge

KlaskTron is an open-hardware physical platform for two-player adversarial reinforcement learning, with a learned actuator model that supports zero-shot sim-to-real transfer of trained policies. Building on this platform, we contribute KlaskTron-IL: a multi-tier imitation learning dataset (80k human demonstration timesteps collected directly on the physical robot, 320k symmetry-augmented timesteps, and synthetic expert trajectories up to 5M steps generated by the zero-shot-validated PPO policy), and diagnostic baselines that expose three characteristic failure modes of standard IL methods under contact-rich adversarial dynamics: BC's data scaling plateau, GAIL's interpretable but weak-gradient discriminator (Spearman ρ=−1.0ρ= -1.0 peg-ball proximity recovery, but a logit span under 0.1), and an asymmetric BC-prior x reward-density interaction. Because the policy interface used throughout this paper is the same one the actuator model bridges to hardware, a policy trained on this benchmark can run unchanged on the physical robot, as we demonstrate zero-shot for the strongest BC baseline, so long as its commands respect the actuator model's ±0.2\pm 0.2 m/s identification range.
Oct 31, 2026

Faster Imitation by Re-timing Demonstrations Within the Controller Limit

Imitation policies inherit the operator's pace, and re-timing the demonstrations to run faster works only until success collapses at two to three times demonstration speed. Recent methods attribute that collapse to distribution shift and answer it with finer rules for which segments may be sped up; we show it has a mechanical cause that segment selection cannot address. A low-level controller realises only a bounded displacement per control period, so merging consecutive commands produces sums outside that bound: training pairs that promise a transition the plant cannot deliver. The share of such pairs in the re-timed data, which we call the exceed share, tracks where the retrained policy collapses. The rule that follows has a single parameter: merge greedily while the sum stays realisable. Every command is then realisable by construction, the demonstrated geometry is preserved, and the speed-up lands only where the demonstrator left slack. On four simulated robomimic tasks, 200 held-out seeds per cell, a merge cap of two is at or above the frozen policy's success on all four, by 14 points on Square, and shortens the successful episodes. At matched tempo, uniform re-timing loses 9 and 16 points on Can and ToolHang, is within noise on the other two, and reaches zero on the hardest at 2x. Placement also matters: no shuffled, entropy-guided or speed-conditioned alternative beats the rule on three of the four tasks, by margins the exceed share alone does not explain; on the fourth, at a tempo the task absorbs, every placement lands in the same band. The same principle carries to a simulated humanoid whole-body tracker: re-timing 429 reference motions within its velocity envelope succeeds on 71% of clips at 2.3x, against 51% for uniform time scaling at 2x.
Oct 30, 2026

FOCUS: First-Order Control Using Optimal Transport Supervision for Imitation from Observations

Imitation learning from observations must recover behavior from state-only demonstrations without known temporal correspondence. Differentiable simulators provide pathwise policy gradients, but differentiating through long, contact-rich trajectories can make optimization unstable. We introduce FOCUS, which combines local optimal-transport alignment with short-horizon differentiable actor-critic learning. FOCUS matches agent rollout windows to multiple sampled expert segments using a soft best-of-KK objective that emphasizes locally compatible matches. The resulting transport-based costs provide differentiable imitation rewards, while a learned critic bootstraps value beyond the rollout horizon. FOCUS requires neither temporally aligned trajectories nor expert-aligned initialization or teacher forcing. Across contact-rich locomotion tasks in DFlex, including a musculoskeletal humanoid controlled by 152 muscle activations, FOCUS improves sample efficiency and training stability over adversarial imitation-from-observation methods and prior differentiable imitation baselines.
Oct 29, 2026

PathCowork: Learning from real-world pathology practice to build AI copilots

Medical AI learns from the answers doctors record, but rarely from how they interpret the evidence to reach them. Here we introduce PathCowork, a framework that turns pathologists' narrated examinations into supervision for visual grounding and procedures that guide diagnostic AI agents. Synchronized gaze, narration and navigation link visual findings to the process of interpreting them. Across 48 sessions from 16 pathologists, the platform captured 1,028 minutes of slide review. Using these data, we trained PathCowork-1T for server deployment and PathCowork-27B for local deployment, with improved visual grounding. Across five organ-specific evaluations, agents combining this grounding with recorded diagnostic procedures achieved primary task scores averaging 17.9 and 24.1 percentage points higher than Claude Code and Codex, respectively, with shorter processing times and lower inference costs. Compared to the state-of-the-art multiple-instance learning AIs, the PathCowork agent-proposed regions achieved higher prostate-grading scores with limited slide labels. PathCowork can also help construct pathology AI benchmarks where questions derived from the recordings identified difficulties in recognizing viewing context and rejecting inapplicable premises. In summary, PathCowork proposes a new paradigm to scale up the data collection and train pathology AI through transparent reasoning from diagnostic sessions, opening a path towards systems that learn directly from routine clinical practice.
Oct 28, 2026

WeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form Video

Producing long-form videos from narration requires interdependent decisions that remain coherent across shots. Existing agentic editors often use multi-stage workflows driven by general models, leaving the quality-cost trade-off of compact, task-adapted policies unclear. We address this gap with WeVisCut, a long-form script-to-video system that connects tool execution with policy learning. Its agent harness maintains shared production state so that tool feedback informs subsequent decisions and repairs. Because finished videos do not expose these decisions, our four-stage data pipeline combines editing inversion with verified replay to turn recovered editing intent into executable supervision. Applied to 3,750 finished videos, this pipeline yields over 130K training examples. We use this supervision to train a single 9B policy that coordinates production and recovers from failures within the same harness. Moreover, we introduce a 100-task benchmark and five metrics assessing editing quality and delivery reliability. Under a shared tool environment, WeVisCut substantially outperforms its base model and achieves comparable performance to Claude Opus 5 while using nearly 33x fewer generated tokens per successful output minute. It also outperforms task-matched native editors on our proposed metrics.
Oct 27, 2026

From 12 to 200 Demonstrations: How ACT, Diffusion Policy, SmolVLA and π0.5π_{0.5} Scale on a Low-Cost Arm

We trained four widely used imitation-learning policies, ACT, Diffusion Policy, SmolVLA and π0.5π_{0.5}, on nested subsets of 12, 25, 50, 100 and 200 teleoperated demonstrations of one pick-and-place task with distractor objects, using one fixed training recipe per policy. Every trained model was then run on a physical SO-101 arm on a fixed set of object placements: 1,340 scored episodes in 126 blocks. ACT, trained from scratch, rose from 1 success in 40 episodes at 12 demonstrations to 37 in 40 at 100. π0.5π_{0.5} with its action expert fine-tuned was the best policy when data was scarce (16 of 40 at 12 demonstrations, 28 of 40 at 25) and then stayed near 70%; from 100 demonstrations ACT was ahead, by a margin this study does not fully resolve. SmolVLA and Diffusion Policy peaked near 50%, and both were affected by a failure to start: SmolVLA's arm never started moving in 47% of episodes, and Diffusion Policy hardly ever started until an idle lead-in was trimmed from the demonstrations. For π0.5π_{0.5} the fine-tuning recipe mattered more than the amount of data: low-rank adaptation succeeded in 4 of 110 episodes, against 130 of 200 for action-expert fine-tuning. From the joint and camera recordings we place every episode on a four-stage ladder (started, reached the ball, carried it, delivered it) and show that for ACT more data moves failures from reaching, to grasping, to none, while the failures π0.5π_{0.5} retains are in grasping. We also measure how weak a ten-episode block is as a measurement (the same checkpoint tested twice differs by 2 successes in 10 on average), separate seed-to-seed variation from that noise, and audit the human scoring against the recordings. Training length was fixed within each recipe, so the amount of data and the number of times it was seen change together; we state which conclusions survive that confound and which do not.
Oct 26, 2026

Closed-Loop Aware Planning via Imitation-Free Speed Scoring for End-to-End Autonomous Driving

End-to-end autonomous driving has predominantly relied on open-loop imitation learning for training, which presents two limitations. First, expert demonstrations show only one behavior per scenario, so the model cannot learn alternative valid behaviors and, in turn, cannot adapt to user-specified preferences at test time. Second, noise or ambiguity in these demonstrations can produce inconsistent, unreliable supervision for the end-to-end models. Driving is inherently multi-modal: multiple different behaviors can be valid responses to the same scene, and this diversity arises largely from speed rather than path. We introduce a planning paradigm that learns the path through imitation while scoring a set of candidate speeds across multiple criteria with a learned function supervised by physically grounded safety constraints. We realize this paradigm as Closed-Loop Aware Planning for Autonomous Driving (CLAP-AD), a plug-in planning head that integrates with existing end-to-end driving models and, to our knowledge, is the first to enable explicit test-time control over their speed selection without speed-conditioned training. On Bench2Drive-Speed, CLAP-AD sets a new state of the art Speed-Adherence Score of 82.53, topping TFv6, the strongest baseline on this metric, by 12.9 points while retaining strong closed-loop driving performance. On Bench2Drive, it improves driving score by +28.5 and +32.7 over UniAD and VAD, and on NAVSIM v2, built on real-world driving data, EPDMS by +6.1 over its backbone. Project page: https://clap-ad.github.io
Oct 26, 2026

Text Is the Bottleneck: A Small-Scale Replication of LLM Game Playing in Riichi Mahjong

We replicate the riichi-mahjong track of "Can Large Language Models Master Complex Card Games?" (arXiv:2509.01328) under a drastically smaller budget: models of at most 2B parameters, one 16 GB consumer GPU, about five GPU-days. (1) With matched data and labels, text is the bottleneck: on the same 100k-decision training set, a 2.6M-parameter convolutional board encoder agrees with the strong AI Mortal on 80.6% of test decisions, above every text-prompt LLM we train (best 77.8%), and places better at the table (2.85 over 1,000 games vs. ~3.0 for the LLMs; paired difference at n=100 not significant). Scaled to 1.71M decisions, a 4.9M-parameter Transformer reaches 94.7% agreement, beats the 0.8B LLMs on identical seeds (-0.45 ranks, sign test p=0.005), and is statistically equivalent to parity against three Mortal champions (TOST within 0.10 ranks, p=6e-4); parity against a teacher-opponent is roughly the ceiling of pure imitation. (2) Training toward the teacher raises teacher agreement more than a 2.5x parameter increase does (+2.5 vs. +1.5 points); on human-record agreement both effects are +1.0, and neither is visible at the table. (3) Few-point agreement differences do not resolve into playing strength within a family: +2.5 agreement points coincide with a paired placement difference of +0.02 +/- 0.12 ranks (p=1.0), and +2.0 points at 2B with -0.06 (p=0.37) - both below the ~0.25-rank resolution of 100 paired games. We also document a paper-code mismatch in the original winner filter, reconstruct its two private engine patches, and show that a divide-and-conquer pipeline makes template-generated Japanese rationales 93% sentence-exact (43% without facts). Per Tenhou's terms we release no trained weights - only code, configurations, and evaluation protocols.
Oct 25, 2026

ElevateHead: Decide Once, Then Commit: When the Action Head Matters for Multimodal Imitation

Robots that learn from people often see several valid ways to do the same task: pass an obstacle on the left or on the right, open the microwave or move the kettle first. A policy that re-decides at every step can waver between these ways mid-task. We ask when the action head (the part of a policy that turns the state into actions) matters, and how it should make this choice. On a controlled task, heads that sample one way succeed far more often than a head that averages the ways when the demonstrations contain several ways and the state gives no hint which to take (+0.72 success with two ways); this trend was not reproduced on the other tasks we ran. We then present ElevateHead, a head that decides once at a learned decision point, commits to that choice for the episode, and exposes the choice as an input a user can set. We designed it by diagnosis: each version was frozen and tested on tasks it had not seen, and the next version changed only what a measured failure pointed to; predictions were registered before each run. The first four versions each fell below BeT on at least one held-out task. The fifth completed all four subtasks of the held-out Franka Kitchen task in 45.2% of episodes against 29.3% for BeT, the strongest baseline across our tasks (+0.160, 95% CI [+0.129, +0.188]); when a mode was forced, the first completed subtask matched it in 95% of successful episodes. Diagnosis also located where commitment breaks: varied start positions (repaired on the controlled task: commitment 0.22 to 0.81, success 0.74 to 0.91) and varied object placement (partly repaired: in an exploratory variant, commitment 0.03 to 0.59 on aligning with no detectable change in success). On held-out block stacking, version 4 was comparable to BeT and did not exceed it. A real-robot test with predictions fixed in advance is planned for a follow-up version of this paper.
Oct 23, 2026

From Freehand Demonstrations to Autonomous Robotic Ultrasound: An Anatomy-Anchored Learning Framework

Ultrasound imaging is highly operator-dependent, motivating robotic systems that reproduce expert scanning skills consistently. These skills involve coordinated interpretation of anatomical landmarks, ultrasound images, and probe-tissue interactions that are difficult to encode in hand-designed control strategies. Imitation learning offers an alternative, but its observation and action representations should reflect the anatomically structured nature of scanning. We propose an anatomy-anchored imitation-learning framework that combines anatomy-guided probe placement with learned ultrasound navigation. The placement module identifies a task-specific anatomical starting point from external visual observations and establishes the initial probe pose as the reference frame for scanning. A flow-based policy then generates scanning motions from anchor-relative probe poses, ultrasound images, probe-mounted camera views, and contact forces, while a termination head predicts task completion. This avoids continuous reliance on third-person vision and reduces dependence on the robot-patient configuration. A handheld acquisition platform enables multimodal demonstrations to be collected during natural freehand scanning without a robot. On 16 unseen participants, task success reaches 96%, 94%, 88%, and 94% for kidney, gallbladder, and liver standard-plane acquisition and radial artery sweeping, respectively. Compared with a fine-tuned pi_0 model, success on the abdominal tasks improves by 10-42 percentage points, with lower time-to-target, peak contact force, and trajectory jerk. Ablations support the complementary roles of anatomical anchoring, egocentric vision, and force feedback. End-to-end trials with automatic placement achieve 90-100% success, and the policies transfer without retraining to a second robot with 83-100% success.
Oct 21, 2026

Externalise the Memory: A Pre-Registered Study of Where an Embodied Agent's Experience Should Live

Should an embodied agent's experience be baked into its policy weights, or kept external and retrieved at inference? A memory-augmented teacher (a VLM that retrieves lessons, with their episode images, from its own past attempts) produces trajectories on a vision-conditioned reach task; from the same trajectories we distil LoRA student policies and evaluate them under matched protocols, including one that re-attaches the teacher's frozen retrieval buffer to a fixed set of student weights. We pre-registered the hypotheses, arms, tests, and analysis rules before training and before any p-value. Distilling the memory-augmented behaviour into weights beat an action-only control by +1 pp (n.s.); the registered test proves underpowered for its own minimum effect, so we withdraw - rather than refute - the memory-in-weights claim. Attaching retrieval to a fixed set of student weights added +34 pp (z=5.15), reaching the pooled memory-augmented teacher reference (49.3%); the pre-registered protocol contrast agrees (+23 pp, z=3.36); under a pre-registered auto-reset sensitivity count the two are +36 and +29 pp. An exploratory behavioural probe with pre-registered per-arm predictions is consistent with behavioural distillation transferring the marginal correction but not the conditional, situation-specific correction that retrieval supplies. The recipe that follows is to externalise the memory: in follow-up controls registered before they ran, the same frozen buffer lifts the base model with no adapter to 48% (sensitivity 48%) and the other three adapters from 19-26% to 53-55% (sensitivity 17-21% to 50-55%), and the adapter adds nothing detectable once the buffer is attached. On a second, harder task, the retrieval advantage points the same direction but falls short of significance (+6 pp, 95% CI [-4.5, +16.4]); exploratory diagnostics tie the attenuation to memory coverage.
Oct 21, 2026

DriveMLP: When Training Strategy Outperforms Model Complexity in Autonomous Driving

Learning motion planning policies that remain reliable under closed-loop execution is a central challenge in Autonomous Driving (AD). Imitation Learning (IL) scales well to large AD datasets but suffers from distribution shift, causal confusion, and compounding errors under closed-loop execution. In contrast, Reinforcement Learning (RL) typically requires handcrafted rule-based rewards and massive simulation. We study a simple closed-loop IL framework that uses RL purely as an optimization tool, guided by a trajectory-distance reward derived from logged expert data. We refer to this training strategy as Expert-Guided RL (EGRL). Through a controlled empirical study we find evidence that the training procedure, rather than model size and complexity, accounts for a substantial part of the performance gains. We then introduce DriveMLP, a lightweight MLP policy trained with EGRL, as a simple instantiation of these findings. It achieves a score of 86.51 on the nuPlan Val14 Reactive benchmark, a +3.2 point improvement over the previous best IL planner (Flow Planner), with >15x faster inference. Code is available at: https://github.com/alessandro-canevaro/DriveMLP .