Human video datasets used for cotraining robot manipulation policies largely consist of curated demonstrations where motions are orchestrated to resemble robot behavior and 3D hand poses are captured with specialized hardware. A more plentiful source of data is everyday Internet video, but it is an open question what factors enable transfer from such videos to robots. We investigate this using a new dataset of 532 human videos with 28 hours of high-quality triangulated hand labels and natural motions. We find that hand pose quality affects transfer, but even with accurate hands, the inherent motion gap hinders transfer unless the vision and policy networks specialize to each embodiment. Our cotraining recipe yields consistent improvements, with an absolute success rate gain of 29.7% in the low-robot-data regime across six manipulation tasks.

Language models must now generalize out of the box to novel environments and work inside inference-scaling search procedures, such as AlphaEvolve, that select rollouts with a variety of task-specific reward functions. Unfortunately, the standard paradigm of LLM post-training optimizes a pre-specified scalar reward, often leading current LLMs to produce low-entropy response distributions and thus to struggle at displaying the diversity that inference-time search will require. We propose Vector Policy Optimization (VPO), an RL algorithm that explicitly trains policies to anticipate diverse downstream reward functions and to produce diverse solutions. VPO exploits that rewards are often vector-valued in practice, like per-test-case correctness in code generation or, say, multiple different user personas or reward models. VPO is essentially a drop-in replacement for the GRPO advantage estimator, but it trains the LLM to output a set of solutions where individual solutions specialize to different trade-offs in the vector reward space. Across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search (e.g. pass@k and best@k), with the gap widening as the search budget grows. For evolutionary search, VPO models unlock problems that GRPO models cannot solve at all. As test-time search becomes more standardized, optimizing for diversity may need to become the default post-training objective.

Reinforcement learning (RL) is increasingly used to post-train vision-language-action (VLA) models, but every update consumes robot rollouts that are slow and costly to collect, making sample efficiency a central concern. Manipulation tasks typically provide only sparse rewards, so a weak policy fails almost every rollout early in training and has little to learn from, even when those failures execute coherent behavior. Such a failure, however, is a success at a different task. We present Learning from Hindsight (LfH), which brings hindsight relabeling to RL post-training of VLAs by scoring failed rollouts against the tasks they actually achieved. A single vision-language model relabels both the instruction and the reward, proposing a hindsight instruction for a group of failed rollouts and scoring how well each satisfies it, and the policy trains on the relabeled and original rollouts jointly. Because VLAs generalize across language, relabeling in language lets the policy learn more from the same trajectories. On out-of-distribution LIBERO-PRO tasks, where standard RL improves only slowly, LfH achieves 5× improvement in sample efficiency, and outperforms a dense progress-reward baseline. The gains hold across VLA backbones and on a physical Franka robot.

Recurrent models process long contexts efficiently by compressing their history into a fixed-size state, but modern architectures typically do so in a single causal pass over the sequence. Each input must therefore be compressed before the model knows how it will later be used, forcing a limited state to compromise across possible future demands. We introduce dynamic compression, which allows a recurrent model to selectively revisit past tokens and revise its fixed-size state through additional recurrent updates. The model need not preserve every part of the history at uniformly high fidelity in its recurrent state, because lower-fidelity information can be revisited from the retained raw sequence when it becomes relevant. We study this in a controlled setting where the model first learns multiple functions in-context and, later in the same sequence, encounters a series of few-shot tasks that each require it to identify and reuse one of those functions. A single-pass model must preserve every function at sufficient fidelity for any future task, whereas selective re-scanning allows the model to revisit and refine only the function currently needed. We find that dynamic compression substantially reduces the recurrent state required for accurate reuse and scales more favorably as the number of stored functions grows. These results demonstrate a computation–memory tradeoff in which recurrent models can spend more computation revisiting their history to make more effective use of a fixed-size state.

Comparison of fine-tuning models with reinforcement learning (RL) and supervised fine-tuning (SFT) reveals that, despite similar performance at a new task, RL preserves prior knowledge and capabilities significantly better. We find that the degree of forgetting is determined by the distributional shift, measured as the KL-divergence between the fine-tuned and base policy evaluated on the new task. Our analysis reveals that on-policy RL is implicitly biased towards KL-minimal solutions among the many that solve the new task, whereas SFT can converge to distributions arbitrarily far from the base model. We validate these findings through experiments with large language models and robotic foundation models and further provide theoretical justification for why on-policy RL updates lead to a smaller KL change. We term this principle $\textit{RL’s Razor}$: among all ways to solve a new task, RL prefers those closest in KL to the original model.

We introduce SoftMimic, a framework for learning compliant whole-body control policies for humanoid robots from example motions. Imitating human motions with reinforcement learning allows humanoids to quickly learn new skills, but existing methods incentivize stiff control that aggressively corrects deviations from a reference motion, leading to brittle and unsafe behavior when the robot encounters unexpected contacts. In contrast, SoftMimic enables robots to respond compliantly to external forces while maintaining balance and posture. Our approach leverages an inverse kinematics solver to generate an augmented dataset of feasible compliant motions, which we use to train a reinforcement learning policy. By rewarding the policy for matching compliant responses rather than rigidly tracking the reference motion, SoftMimic learns to absorb disturbances and generalize to varied tasks from a single motion clip. We validate our method through simulations and real-world experiments, demonstrating safe and effective interaction with the environment.

Pre-training is crucial for large language models (LLMs), as it is when most representations and capabilities are acquired. However, natural language pre-training has problems: high-quality text is finite, it contains human biases, and it entangles knowledge with reasoning. This raises a fundamental question: is natural language the only path to intelligence? We propose using neural cellular automata (NCA) to generate synthetic, non-linguistic data for pre-pre-training LLMs–training on synthetic-then-natural language. NCA data exhibits rich spatiotemporal structure and statistics resembling natural language while being controllable and cheap to generate at scale. We find that pre-pre-training on only 164M NCA tokens improves downstream language modeling by up to 6

Exploration is essential to RL since a policy cannot improve by repeatedly sampling the behaviors it already prefers. Standard methods inject stochasticity in the action space, but such jitter only yields rollouts close to the original. Escaping a weak policy often requires global perturbations that action noise cannot produce. Large language models (LLMs) and vision-language-action (VLA) models offer a pathway: they condition the policy on a natural language prompt, and since the rollout follows from it, modifying the prompt induces global changes. The challenge is finding prompts that induce useful global changes. With a weak policy that rarely succeeds, reward is too sparse to select on. Our idea is to refine prompts from the rollouts themselves: a vision-language model (VLM) reasons over the rollout video, diagnoses how the policy responded, and rewrites the prompt to elicit better behavior next time. This procedure realizes posterior sampling, a classical RL exploration framework, at the level of prompts: the VLM maintains an implicit distribution over useful prompts and updates it from observed rollouts. We call this strategy Prompt-Driven Exploration (PDE). Across manipulation and reasoning tasks, PDE enables RL to learn successful policies even from zero-reward starts, and improves sample efficiency more broadly. Our website is available at https://xinyunsunshine.github.io/prompt-rl.

Traditional robotic hand metrics focus on static properties such as workspace, manipulability, and grasp stability. However, these metrics do not directly measure dexterity under the standard definition in robotic manipulation: the ability to continuously change an object's pose within the hand while maintaining contact from an initial grasp. We introduce Kinematic Rolling Manipulation Ability (KaRMA), a kinematic-only metric for fine manipulation that quantifies reachable in-hand translation and reorientation of a spherical test object within a two-finger precision pinch through feasible rolling motions. KaRMA enforces joint limits, collision constraints, rolling contact, and antipodal force feasibility, then investigates reachable in-hand object poses via breadth-first search over translation and rotation primitives. KaRMA reports three scores: translational coverage (KaRMA-T), rotational coverage (KaRMA-R), and sensitivity to the initial grasp (KaRMA-S). We evaluate KaRMA on 16 widely used robotic hands and compare against static baselines, showing that KaRMA separates hands that rank identically under static proxies, reveals translation-rotation tradeoffs invisible to existing baselines, and is qualitatively consistent with selected published task benchmarks where Jacobian-based metrics can be misleading.

Position controllers have become the dominant interface for executing learned manipulation policies. Yet a critical design decision remains understudied: how should we choose controller gains for policy learning? The conventional wisdom is to select gains based on desired task compliance or stiffness. However, this logic breaks down when controllers are paired with state-conditioned policies: effective stiffness emerges from the interplay between learned reactions and control dynamics, not from gains alone. We argue that gain selection should instead be guided by learnability: how amenable different gain settings are to the learning algorithm in use. In this work, we systematically investigate how position controller gains affect three core components of modern robot learning pipelines: behavior cloning, reinforcement learning from scratch, and sim-to-real transfer. Through extensive experiments across multiple tasks and robot embodiments, we find that: (1) behavior cloning benefits from compliant and overdamped gain regimes, (2) reinforcement learning can succeed across all gain regimes given compatible hyperparameter tuning, and (3) sim-to-real transfer is harmed by stiff and overdamped gain regimes. These findings reveal that optimal gain selection depends not on the desired task behavior, but on the learning paradigm employed. Project website: https://younghyopark.me/tune-to-learn

Continual learning, enabling models to acquire new skills and knowledge without degrading existing capabilities, remains a fundamental challenge for foundation models. While on-policy reinforcement learning can reduce forgetting, it requires explicit reward functions that are often unavailable. Learning from expert demonstrations, the primary alternative, is dominated by supervised fine-tuning (SFT), which is inherently off-policy. We introduce Self-Distillation Fine-Tuning (SDFT), a simple method that enables on-policy learning directly from demonstrations. SDFT leverages in-context learning by using a demonstration-conditioned model as its own teacher, generating on-policy training signals that preserve prior capabilities while acquiring new skills. Across skill learning and knowledge acquisition tasks, SDFT consistently outperforms SFT, achieving higher new-task accuracy while substantially reducing catastrophic forgetting. In sequential learning experiments, SDFT enables a single model to accumulate multiple skills over time without performance regression, establishing on-policy distillation as a practical path to continual learning from demonstrations.

Legged robot research is presently focused on bipedal or quadrupedal robots, despite capabilities to build robots with many more legs to potentially improve locomotion performance. This imbalance is not necessarily due to hardware limitations, but rather to the absence of principled control frameworks that explain when and how additional legs improve locomotion performance. In multi-legged systems, coordinating many simultaneous contacts introduces a severe curse of dimensionality that challenges existing modeling and control approaches. As an alternative, multi-legged robots are typically controlled using low-dimensional gaits originally developed for bipeds or quadrupeds. These strategies fail to exploit the new symmetries and control opportunities that emerge in higher-dimensional systems. In this work, we develop a principled framework for discovering new control structures in multi-legged locomotion. We use geometric mechanics to reduce contact-rich locomotion planning to a graph optimization problem, and propose a spin model duality framework from statistical mechanics to exploit symmetry breaking and guide optimal gait reorganization. Using this approach, we identify an asymmetric locomotion strategy for a hexapod robot that achieves a forward speed of 0.61 body lengths per cycle (a 50

Large Language Models (LLMs) have demonstrated impressive real-world utility, exemplifying artificial useful intelligence (AUI). However, their ability to reason adaptively and robustly – the hallmarks of artificial general intelligence (AGI) – remains fragile. While LLMs seemingly succeed in commonsense reasoning, programming, and mathematics, they struggle to generalize algorithmic understanding across novel contexts. Our experiments with algorithmic tasks in esoteric programming languages reveal that LLM's reasoning overfits to the training data and is limited in its transferability. We hypothesize that the core issue underlying such limited transferability is the coupling of reasoning and knowledge in LLMs. To transition from AUI to AGI, we propose disentangling knowledge and reasoning through three key directions: (1) pretaining to reason using RL from scratch as an alternative to the widely used next-token prediction pretraining, (2) using a curriculum of synthetic tasks to ease the learning of a reasoning prior for RL that can then be transferred to natural language tasks, and (3) learning more generalizable reasoning functions using a small context window to reduce exploiting spurious correlations between tokens. Such a reasoning system coupled with a trained retrieval system and a large external memory bank as a knowledge store can overcome several limitations of existing architectures at learning to reason in novel scenarios.

The scarcity of diverse and high-quality data impedes the quest to build a generalist robotic system. Current robotics data collection efforts face many challenges: the need for physical robotic hardware, setting up the environment, frequent resets, and the fatigue for data collectors operating real robots. We introduce DART, a teleoperation platform designed for crowdsourcing that reimagines robotic data collection by leveraging cloud-based simulation and augmented reality (AR) to address many limitations of prior data collection efforts. User studies show that DART enables higher data collection throughput and lower physical fatigue than real-world teleoperation. We also demonstrate that policies trained using DART-collected datasets successfully transfer to reality and are robust to unseen visual disturbances. All data collected through DART is automatically stored in a cloud-hosted database, DexHub, paving the path for an ever-growing data hub for robot learning. https://dexhub.ai/project

Recent advancements in state-of-the-art (SOTA) offline reinforcement learning (RL) have primarily focused on addressing function approximation errors, which contribute to the overestimation of Q-values for out-of-distribution actions, a challenge that static datasets exacerbate. However, high stakes applications such as recommendation systems in online gaming, introduce further complexities due to player's psychology (intent) driven by gameplay experiences and the inherent volatility on the platform. These factors create highly sparse, partially overlapping state spaces across policies, further influenced by the experiment path selection logic which biases state spaces towards specific policies. Current SOTA methods constrain learning from such offline data by clipping known counterfactual actions as out-of-distribution due to poor generalization across unobserved states. Further aggravating conservative Q-learning and necessitating more online exploration. FAST-Q introduces a novel approach that (1) leverages Gradient Reversal Learning to construct balanced state representations, regularizing the policy-specific bias between the player's state and action thereby enabling counterfactual estimation; (2) supports offline counterfactual exploration in parallel with static data exploitation; and (3) proposes a Q-value decomposition strategy for multi-objective optimization, facilitating explainable recommendations over short and long-term objectives. These innovations demonstrate superiority of FAST-Q over prior SOTA approaches and demonstrates at least 0.15 percent increase in player returns, 2 percent improvement in lifetime value (LTV), 0.4 percent enhancement in the recommendation driven engagement, 2 percent improvement in the player's platform dwell time and an impressive 10 percent reduction in the costs associated with the recommendation, on our volatile gaming platform.

Animals and robots navigate through environments by building and refining maps of space. These maps enable functions including navigation back to home, planning, search and foraging. Here, we use observations from neuroscience, specifically the observed fragmentation of grid cell map in compartmentalized spaces, to propose and apply the concept of Fragmentation-and-Recall (FARMap) in the mapping of large spaces. Agents solve the mapping problem by building local maps via a surprisal-based clustering of space, which they use to set subgoals for spatial exploration. Agents build and use a local map to predict their observations; high surprisal leads to a "fragmentation event" that truncates the local map. At these events, the recent local map is placed into long-term memory (LTM) and a different local map is initialized. If observations at a fracture point match observations in one of the stored local maps, that map is recalled (and thus reused) from LTM. The fragmentation points induce a natural online clustering of the larger space, forming a set of intrinsic potential subgoals that are stored in LTM as a topological graph. Agents choose their next subgoal from the set of near and far potential subgoals from within the current local map or LTM, respectively. Thus, local maps guide exploration locally, while LTM promotes global exploration. We evaluate FARMap on complex procedurally-generated spatial environments and realistic simulations to demonstrate that this mapping strategy much more rapidly covers the environment (number of agent steps and wall clock time) and is more efficient in active memory usage, without loss of performance.

We introduce two multilingual, multimodal foundation language models that power Apple Intelligence features across Apple devices and services: i a 3B-parameter on-device model optimized for Apple silicon through architectural innovations such as KV-cache sharing and 2-bit quantization-aware training; and ii a scalable server model built on a novel Parallel-Track Mixture-of-Experts PT-MoE transformer that combines track parallelism, mixture-of-experts sparse computation, and interleaved global-local attention to deliver high quality with competitive cost on Apple's Private Cloud Compute platform. Both models are trained on large-scale multilingual and multimodal datasets sourced via responsible web crawling, licensed corpora, and high-quality synthetic data, then further refined with supervised fine-tuning and reinforcement learning on a new asynchronous platform. The resulting models support several additional languages while understanding images and executing tool calls. In public benchmarks and human evaluations, both the server model and the on-device model match or surpass comparably sized open baselines. A new Swift-centric Foundation Models framework exposes guided generation, constrained tool calling, and LoRA adapter fine-tuning, allowing developers to integrate these capabilities with a few lines of code. The latest advancements in Apple Intelligence models are grounded in our Responsible AI approach with safeguards like content filtering and locale-specific evaluation, as well as our commitment to protecting our users' privacy with innovations like Private Cloud Compute.

Recent advances in Behavior Cloning (BC) have made it easy to teach robots new tasks. However, we find that the ease of teaching comes at the cost of unreliable performance that saturates with increasing data for tasks requiring precision. The performance saturation can be attributed to two critical factors: (a) distribution shift resulting from the use of offline data and (b) the lack of closed-loop corrective control caused by action chucking (predicting a set of future actions executed open-loop) critical for BC performance. Our key insight is that by predicting action chunks, BC policies function more like trajectory “planners” than closedloop controllers necessary for reliable execution. To address these challenges, we devise a simple yet effective method, Resip (Residual for Precise Manipulation), that overcomes the reliability problem while retaining BC's ease of teaching and long-horizon capabilities. Resip augments a frozen, chunked BC model with a fully closed-loop residual policy trained with reinforcement learning (RL) that addresses distribution shifts and introduces closed-loop corrections over open-loop execution of action chunks predicted by the BC trajectory planner. Videos, code, and data: residual-assembly.github.io.

Background: Major Adverse Kidney Events within 30 days (MAKE30) is an important patient-centered outcome for assessing the impact of acute kidney injury (AKI). The existing prediction models for MAKE30 are static and overlook dynamic changes in clinical status. In this study, we introduce ORAKLE, a novel deep-learning model that utilizes evolving time-series data to predict MAKE30, enabling personalized, patient-centered approaches to AKI management and outcome improvement. Methods: We conducted a retrospective study using three publicly available critical care databases: MIMIC-IV, SICdb, and eICU-CRD. Among these, MIMIC-IV was divided into 80% training and 20% internal test sets, whereas SiCdb and eICU-CRD were used as external validation cohorts. Patients with sepsis-3 criteria who developed AKI within 48 hours of intensive care unit admission were identified. Our primary outcome was MAKE30, defined as a composite of death, new dialysis or persistent kidney dysfunction within 30 days of ICU admission. We developed ORAKLE using Dynamic DeepHit framework for time-series survival analysis and its performance against Cox models using AUROC and AUPRC. We further assessed model calibration using Brier score. Results: We analyzed 16,671 patients from MIMIC-IV, 2,665 from SICdb, and 11,447 from eICU-CRD. ORAKLE outperformed the Cox models in predicting MAKE30, achieving AUROCs of 0.84 (95% CI: 0.83-0.86) vs. in MIMIC-IV internal test set 0.80 (95% CI: 0.78-0.82), 0.83 (95% CI: 0.81-0.85) vs. 0.79 (95% CI: 0.77-0.81) in SICdb, and 0.85 (95% CI: 0.84-0.85) vs. 0.81 (95% CI: 0.80-0.82) in eICU-CRD. The AUPRC values for ORAKLE were also significantly better than that of Cox models. The Brier score for ORAKLE was 0.21 across the internal test set, SICdb, and eICU-CRD, suggesting good calibration. Conclusions: ORAKLE is a robust deep-learning model for predicting MAKE30 in critically ill patients with AKI that utilizes evolving time series data. By incorporating dynamically changing time series features, the model captures the evolving nature of kidney injury, treatment effects, and patient trajectories more accurately. This innovation facilitates tailored risk assessments and identifies varying treatment responses, laying the groundwork for more personalized and effective management approaches. ### Competing Interest Statement GNN is a founder of Renalytix, Pensieve, Verici and provides consultancy services to AstraZeneca, Reata, Renalytix, Siemens Healthineer and Variant Bio, serves a scientific advisory board member for Renalytix and Pensieve. He also has equity in Renalytix, Pensieve and Verici. JAK reports receiving consulting fees from Astute Medical/bioMerieux, Astellas, Alexion, Chugai Pharma, Novartis, Mitsubishi Tenabe and GE Healthcare and is a Full-time employee of Spectral Medical. All remaining authors have declared no conflicts of interest. ### Funding Statement This study was supported by NIH grant K08DK131286 (AS). The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: The study received approval from the Institutional Review Board at the Icahn School of Medicine at Mount Sinai (approval no. 19-00951). I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes Publicly available datasets were analyzed in this study. The MIMIC-IV dataset is available at https://physionet.org/content/mimiciv/, the eICU-CRD dataset is available at https://physionet.org/content/eicu-crd/, and the SICdb dataset is available at https://physionet.org/content/sicdb/.

Scaling robot learning requires data collection pipelines that scale favorably with human effort. In this work, we propose Crowdsourcing and Amortizing Human Effort for Real-to-Sim-to-Real(CASHER), a pipeline for scaling up data collection and learning in simulation where the performance scales superlinearly with human effort. The key idea is to crowdsource digital twins of real-world scenes using 3D reconstruction and collect large-scale data in simulation, rather than the real-world. Data collection in simulation is initially driven by RL, bootstrapped with human demonstrations. As the training of a generalist policy progresses across environments, its generalization capabilities can be used to replace human effort with model generated demonstrations. This results in a pipeline where behavioral data is collected in simulation with continually reducing human effort. We show that CASHER demonstrates zero-shot and few-shot scaling laws on three real-world tasks across diverse scenarios. We show that CASHER enables fine-tuning of pre-trained policies to a target scenario using a video scan without any additional human effort. See our project website: https://casher-robot-learning.github.io/CASHER/