Autonomous agents must continually acquire reusable knowledge from experience without relying on costly large language model (LLM) reasoning for every decision. However, existing experience-based approaches often struggle to transform accumulated observations and reasoning into compact knowledge that lightweight models can independently apply. This paper proposes SCALE, a small-model-first framework that integrates selective LLM reasoning, experience-driven concept abstraction, and knowledge internalization. A lightweight model handles routine predictions, while prediction errors and accumulated cross-episode evidence selectively trigger LLM reflection. The LLM analyzes relevant historical experiences, generates positive and negative counterfactual examples, and proposes reusable concept hypotheses. Independently validated concepts are progressively incorporated into hierarchical neural representations, enabling lightweight models to exploit increasingly abstract knowledge. Meanwhile, mastered experiences are excluded from active memory while their original records remain available in a complete archive. Experiments across ten random seeds show that historical recall improves held-out accuracy from 87.11% to 98.28%, hierarchical concept knowledge increases difficult-case accuracy from 55.33% to 66.47%, and abstract archiving reduces active experience records from 6,000 to 1,528.
Large-scale representation learning has produced public foundation models that provide strong general-purpose visual features. However, their pretraining objectives are not designed for robot representation learning, so the resulting embeddings tend to emphasize global semantics rather than the temporally sensitive representations required for robot control. Training robot models from scratch is also undesirable due to the limited scale and high collection cost of robotics data, as well as the substantial computational burden of pre-training. In this paper, we propose SPARK (Simple Post-training for Adapting pRetrained Knowledge), a post-training method that adapts foundation models to robot control. SPARK is built on two principles: temporal abstraction, which encourages the encoder to derive compact features that preserve information essential for understanding temporal changes and action-relevant scene structure, and knowledge preservation, which aligns patch-level representations with those of the original foundation model to retain useful pretrained semantics. These objectives yield compact visual state representations that remain semantically meaningful while preserving useful prior knowledge. Experiments across multiple robotics benchmarks show that SPARK consistently improves success rates and generalization over vanilla foundation models, and further demonstrate that these gains transfer to real-world robot manipulation.
Embodied navigation requires agents to achieve persistent spatial awareness and long-horizon execution stability in open-world environments. However, existing navigation architectures struggle to sustain both capabilities, remaining effective only over a handful of recent frames. We present Compass 1, the first embodied navigation foundation model to unify language understanding and video generation while enabling memory backtracking through nearly 2mins of visual history. Without altering the underlying architecture, we introduce History Representation, which integrates sparse visual keyframes with an explicit History Representation Map to preserve long-term spatiotemporal context. To effectively leverage this history, we further introduce History Utilization, a structured VQA supervision protocol comprising three categories: task-stage grounding, exploration consistency, and spatiotemporal target retrieval, enabling map-grounded reasoning and stable persistent long-horizon navigation. To fuel this system, we establish a complete annotation pipeline for monocular RGB videos and open-source a massive dataset featuring over 500 hours of high-quality real-world interactions and over 2,000 hours of annotated web videos. Trained mainly on the distribution of 1m height camera, Compass 1 generalizes seamlessly across diverse embodiments and scenarios, ranging from indoor environments to pedestrian-dense streets, spanning wheeled robot at 0.35m, quadruped dog at 0.5m, humanoid at 0.85m, and aerial drone at 1.5m. Compass 1achieves a dominant overall success rate of 79.4% across open-world evaluations, compared to 17.8% for the strongest baseline. It demonstrates exceptional zero-shot execution exceeding 90% on foundational tasks, a 250% boost in exploration, and an unparalleled leap from 0% to 58% in challenging persistent history backtracking.
Dexterous grasping in robotics relies on distinct hand configurations depending on object geometry. We introduce BetaGrasp, a 3D diffusion policy that conditions a DP3-style point-cloud policy on a four-class grasp label (crimp, sloper, pinch, and jug), running on a Franka arm with a 16-DoF LEAP hand. We also introduce a climbing-hold benchmark: a spring testbed whose ratchet gives a slip-force reading in newtons. In our analysis the policy's point-cloud encoder did not separate the four holds, so the scan alone gave it little to choose a grasp from. Without the label (a DP3-style baseline), simulated rollouts from the robot's reset pose all end in a jug grasp, whatever the hold (25% correct grasp type, versus 94% for BetaGrasp). The label works as a switch for which grasp the policy performs. On 80 paired hardware trials, BetaGrasp raises average slip force from 9.6 N to 14.3 N, and grasps slipping fully off the hold drop from 44% of trials to 10%.
Recent Vision-Language-Action (VLA) models have begun incorporating thermal sensing, but existing approaches typically treat thermal images as ordinary visual inputs, providing limited support for temperature-aware relational reasoning in manipulation. To address this limitation, we propose Thermo-VLA with a ThermoRes Encoder, which decomposes thermal images into cold and hot attribute representations and injects them into the RGB token through language-conditioned residual modulation, enabling temperature-driven target selection and continuous manipulation. We further construct six real-world robotic tasks on a dual-arm Unitree G1 platform, spanning basic grounding to relational reasoning. Experiments show that Thermo-VLA achieves an average success rate of 0.86 across four thermal-dependent tasks, outperforming a pi_0.5 baseline that directly incorporates RGB-thermal observations by 18 percentage points. Under unseen-object and scene variations, Thermo-VLA maintains an average success rate of 0.78, achieving the best generalisation performance among the compared methods. Beyond thermal sensing, we transfer the same residual modulation mechanism to colour attributes and observe consistent improvements in both simulation and real-robot experiments, indicating its potential to generalise beyond thermal attributes. Project website: https://j-x-chen.github.io/Thermo-VLA-web/
Robot demonstration datasets combine demonstrations from teleoperators of uneven experience, so collection teams need to know how many demonstrations from one group match those from another. We measure this with a data multiple: the ratio of the numbers of demonstrations two groups need for their policies to reach the same success rate. On robomimic's simulated Square and Can tasks, we train behavior-cloning policies on demonstrations from the two teleoperators robomimic labels Better and the two it labels Worse, varying dataset size and the share of Worse demonstrations. On Square, within robomimic's 25-second time limit, an all-Worse training set needs 1.55 times as many demonstrations as an all-Better set. Most of this difference is late completion: policies trained on Worse demonstrations finish later, as the Worse teleoperators do, and with a 100-second limit the multiple falls to 1.11. On Can, most of the difference remains at 100 seconds. If each Worse demonstration were worth 1/1.55 of a Better one, as in an all-Worse set, a half-Better, half-Worse set on Square would need 1.22 times as many demonstrations as an all-Better set. In fact, it needs 13% more than that, and this excess persists at every time limit from 20 to 200 seconds. The excess remains when training draws demonstrations uniformly instead of in proportion to their length. On Square, within the 25-second limit, switching to these Worse teleoperators is cheaper at equal success only if they earn less than a third of the Better teleoperators' hourly pay, because Worse demonstrations also take longer to record.
The wide field of view of fisheye cameras and the simplicity of a monocular setup make monocular fisheye visual odometry (VO) attractive for robotic localization and mapping. Nevertheless, scale drift remains a major challenge over long trajectories. Unsupervised depth learning can help to mitigate this issue, but the pronounced distortion of fisheye images poses challenges for accurate depth estimation. This article presents S$^2$D-FVO, a monocular fisheye VO framework integrating spherical bundle adjustment (SBA)-guided unsupervised depth learning to improve scale consistency and depth estimation accuracy. Given that pixel reprojection residuals do not uniformly reflect angular deviations between viewing rays and that the first-order linearization of a scalar angular reprojection residual is insensitive to perturbations along one tangent direction, SBA employs tangent-space angular reprojection residuals derived from the spherical logarithmic map. These residuals retain sensitivity along both tangent directions, with their norms equal to the corresponding scalar angular residuals. A median depth-based scale regularization term is further integrated into SBA to constrain the reconstruction scale. Building on these formulations, SBA replaces the pose estimation network during unsupervised depth learning, providing more reliable relative poses for view synthesis. In turn, depth estimates from the well-trained depth estimation network provide scale references for SBA, improving scale consistency while allowing individual 3-D landmark positions to be refined. Extensive experiments conducted on multiple public datasets, as well as two newly collected large-scale datasets, demonstrate that the proposed framework outperforms representative existing approaches in both depth and trajectory estimation accuracy and is applicable to different camera models.
Generalist robot policies often fail at one step of a task they can otherwise perform. We ask whether such a failure can be repaired by editing the policy's weights directly, without gradients, training data, or a change to the instruction. Told to put the cube then the banana in the bowl, a zero-shot $\pi_{0.5}$ vision-language-action policy in RoboLab releases the cube before it reaches the bowl. Replaying 21 of these decisions under rewordings shows that it binds "in the bowl" to the wrong object. Closed-form instruction distillation copies the per-layer effect of an explicit rewording into the language model's MLP weights. On 128 held-out seeds, success rises from 27 to 106 episodes, with 81 fixed and 2 broken ($p=7\times10^{-22}$), while an update of the same size in random directions leaves success unchanged. The edit also lifts two other tasks without detectable harm on the remaining 13, and a variant that stores 0.16% of the parameter count keeps most of its effect. Placed in the action expert or action head instead, the same distillation helps less or not at all. Our released tool reproduced the fix end to end on new held-out seeds, from 8/64 to 36/64. Edits aimed at perception, a missing grasp, contact physics, or the gripper threshold did not help, and of seven blind Claude Opus editors, only the one given the misread instruction produced an edit that passed held-out scoring. These results suggest that weight edits work best when a known-good input already exists to distil. Because reruns with unchanged weights flip about a third of paired outcomes, evaluations should pair seeds, rerun the unedited policy, and confirm gains on held-out seeds.
One-step stochastic world models avoid the mode averaging of deterministic feature predictors without the tens of network evaluations per step that diffusion and flow matching require. The simplest such objective is nearest-of-m matching (conditional implicit maximum likelihood estimation, IMLE): draw m candidate futures and pull only the one closest to the observed outcome. Recent work conjectures that a small m suffices for feature-space world models and that a strictly proper energy score should replace the hard minimum. We test both claims with one-step world models on frozen DINOv2/v3 patch tokens, using benchmarks whose futures can be enumerated. Under mode-like uncertainty, nearest-of-m with m=16 covers the enumerated futures better than 50-step flow matching (at about 1/30 of its latency), one-step flow generators and an energy-score generator at matched parameter counts, and its rollouts stay on the token manifold where flow rollouts drift. One quantity governs the choice of m: the nearest-of-m objective is the expected distortion of a random codebook of m draws from the generator. This view yields three predictions, recorded before the corresponding runs: (i) memorization of the realized outcome in contexts that never repeat, whose strength is governed by the codebook's miss distortion I(m) rather than its miss probability, and which an energy-score generator does not escape by raising m; (ii) a flattening of mode probabilities with m, matching a known optimum within 0.01; and (iii) under-dispersion on continuous high-dimensional outcomes, which bounds the method on real robot video. Per-step matching on self-rollouts cuts long-horizon error by a third on a stochastic Push-T variant; an ablation separates its manifold and coverage benefits. Planning success does not improve despite accurate risk estimates, a negative result we report with its diagnosis.
Fleets of driving vision-language-action (VLA) models are developed one domain at a time: each operator finetunes a shared base model on private data, producing specialized experts that are costly to consolidate into a single model. Weight merging loses accuracy as heterogeneous domains accumulate, and dense mixture-of-experts routing is prohibitively expensive to combine them. We present DeltaForge, to our knowledge the first post-hoc mixture-of-experts VLA for driving, which consolidates N independently finetuned per-domain models into a single multi-domain model from their weights and a small proxy set alone. When different operators finetune the same base model, their weight updates end up sharing a lot of structure. We exploit this by decomposing all the updates together using singular value decomposition (SVD), splitting each one into a shared part (the same across all domains) and a tiny domain-specific part. The router then only needs to work with those tiny domain-specific pieces, not the full expert weights, which is what keeps the approach efficient. Then two cooperating clip-level blind routers, one over the VLM reasoning experts and one over the flow-matching dynamics experts, are trained from soft labels on the proxy dataset. Across six diverse domains spanning urban, off-road, and highway driving on cars, trucks, and all-terrain vehicles (ATVs), DeltaForge approaches the accuracy of the per-domain experts and surpasses every merging, routing, centralized, and federated baseline at close to single-model parameter and compute cost. Adding a new domain requires only projecting its weight update onto the shared basis with no end-to-end retraining needed. The resulting model also generalizes to domain combinations it was never explicitly trained on.
Connected physical systems fuse observations from robots, IoT sensors, and edge-AI modules into a shared account of the world, yet most evaluations score only the final answer or action and cannot tell whether a method attributed the evidence, updated its belief, handled contradictions, and justified its decision. We argue that shared reasoning in Connected Robotics should be evaluated as an evidence-to-belief trajectory, and introduce the Connected Robotics Evidence Graph (CREG) Benchmark to make that trajectory measurable. CREG comprises 350 deterministically generated scenarios across seven regimes (justified revision, belief preservation, relation association, contradiction onset and resolution, stale/delayed evidence, and source reliability), each exposing a normative posterior reference at every evidence prefix, validated by an independent log-odds checker and screened by a blinded expert audit. Nine representation conditions separate relation content from serialization, ordering, and metadata. Evaluating four open-weight instruction models and six transparent fusion baselines yields three findings. First, entity grounding accounts for a substantial part of the weak-model deficit. Second, deleting semantic relations costs 15.3 points of strict belief-state accuracy on relation-dependent scenarios (Holm-corrected p<0.001), but an information-matched structured list outperforms the full graph and shuffling the graph does not hurt: relation content matters where required, while graph serialization confers no consistent advantage. Third, belief accuracy is strongly non-monotonic along the evidence sequence, with mid-trajectory failures that endpoint-only evaluation cannot detect. All scenarios, references, prompts, metrics, and analysis code are released for offline, hardware-free reproduction.
3D visual perception and scene reconstruction are critical for autonomous robots and intelligent transportation, yet real-world deployments often generate long, redundant image sequences that hinder downstream processing. This report investigates BCVS (budget-aware correspondence-driven view sampling), a pose-free view selection algorithm that offers 22+ FPS on a CPU, enabling online filtering for efficient 3D perception by mobile robots. Unlike existing approaches that rely on full structure-from-motion (SfM) pipelines or camera pose estimation, BCVS employs a purely correspondence-based scoring strategy that prioritizes view novelty while maintaining sufficient overlap for geometric stability. Given a user-specified budget, BCVS iteratively selects the most distinct frames and suppresses near-duplicates, producing compact subsets that preserve connectivity and minimize reconstruction loss. Extensive experiments on benchmark show that BCVS outperforms CPU-friendly baselines at tight budgets and remains competitive as budgets scale. Field trials with large aerial and marine robotics data further validate its practicality and scalability: BCVS reduces redundancy up to 16% while sustaining robust SfM registration and high-quality 3D reconstructions.
Robots are typically evaluated in episodes that begin from a reset, while industrial work consists of repeated jobs whose outcomes affect subsequent operations. Evaluating such work at scale requires simulation of task-critical physical interactions, construction of industrial environments, and measurements of performance as scene state persists. We present HOLIDAYFACTORY 1.0, a framework for simulating, generating, and evaluating industrial robot work. Given a human-specified workplace and task, the framework defines a world pairing a scenario of scene variants and process logic with an evaluation defining a job, metrics, and a rubric. HOLIDAYFACTORY 1.0 comprises three components. HOLIDAY-SIM extends Newton's SDF-based hydroelastic contact with additional constitutive laws and a semi-implicit primal-dual solver, couples stateful actuator dynamics to articulated motion, and derives tactile and joint observations from simulated states. Agentic world construction generates simulation-ready assets and scene variants using geometric and interaction checks, and proposes task-conditioned evaluations for human confirmation and executable checks. 24/7-BENCH introduces seven industrial tasks in which jobs are evaluated without resetting the scene, measuring completed, failed, and unfinished jobs, throughput, latency, and motor energy. Controlled studies characterize the proposed contact, actuator, and tactile models and examine the effectiveness of verification-guided world construction. Experiments on scene and evaluation generation further reveal the limitations of model critics and the difficulty of ensuring that generated evaluations faithfully capture task requirements. We examine job-level evaluations with scripted trajectories (e.g., code-as-policy) but have not yet compared policies over persistent runs.
Sampling-based planners such as MPPI and CEM query a learned world model thousands of times per control step, so low-bit world models are an obvious way to make model-based control cheaper on a robot. With round-to-nearest activation quantization and no calibration, we find that planning can collapse, by up to 100 points of success, while the world model's latent one-step prediction error stays no larger than that of a stochastic rounding scheme that plans well. At the same bit width, what decides the outcome is not the size of the rounding error but whether it persists across planner iterations and is shared across candidates. Round-to-nearest leaves an error that persists, and an iterating planner drifts toward the candidates that this error favors; in TD-MPC2, holding the same dither fixed within a step or an episode brings the drift back. Dither shared by all candidates and redrawn at every iteration removes most of the damage at the same bit width and without training: at least half of it in all 19 affected Meta-World cells (task $\times$ bit width) of TD-MPC2 (median 95%), in 4 of 5 DINO-WM cells, and in all 8 affected tasks of a 317M-parameter multi-task TD-MPC2. Captured as CUDA graphs on an Ampere GPU, a W8A4 integer TD-MPC2 world model runs its planning rollouts 1.9-3.3$\times$ and the whole planning step 1.3-1.5$\times$ faster than FP32, though not faster than BF16 with pre-converted weights; shared dither adds at most 2.3% latency. In TD-MPC2, block-scaled NVFP4 removes most of the damage and MXFP4 part of it, which ties the failure to coarse grids with a shared scale. Hypotheses, thresholds and analysis code were fixed before the data that test them, and the hypotheses that failed are reported.
Generative models can support both state estimation and control in robotics. For control, a common approach is to use the model as a policy that directly samples actions conditioned on observation histories. However, models that jointly represent states, actions, and observations can also predict how the robot state and its environment evolve. We investigate whether these predictions can improve state estimation and control through a change in inference procedure, without modifying the pretrained model. To this end, we use particle filtering to maintain competing hypotheses about state histories. A single pretrained diffusion model generates these hypotheses along with their predicted observations and associated action proposals. As new measurements arrive, particles are reweighted according to how well their predicted observations match the measurements. While particle filtering conventionally uses these weights only for state estimation, we also use them to guide action selection, favoring proposals associated with hypotheses that better explain the observations. Simulation experiments in partially observed navigation and vision-based manipulation demonstrate improvements in both state estimation and task performance compared to action generation without filtering, using the same pretrained model and no additional test-time information.
Reinforcement learning (RL) requires large amounts of interaction data or a reliable simulator, both hard to obtain for nonlinear soft robotic hardware. Recent work shows that a time-scale separation inherent to soft robot dynamics justifies their reduction to low-dimensional invariant manifolds, known as adiabatic spectral submanifolds (aSSMs). We demonstrate that an aSSM-reduced model, a simple polynomial ordinary differential equation learned purely from pre-collected data, can serve as an RL training environment. On a pneumatically actuated soft arm, we train a trajectory-tracking policy in 3 hours on a six-dimensional aSSM-reduced model and deploy it without fine-tuning, whereas reproducing this policy training extrapolates to 3 years of hardware interaction. Our policy tracks dynamic targets with 40% lower error than otherwise identical policies trained on a 4.8-million-parameter recurrent neural network (RNN) or a 1.7-million-parameter transformer, which take 13 and 100 hours to identify, against 5 minutes for our mathematically justified 6,924-parameter model.
A purely data-driven method grounded in the theory of adiabatic spectral submanifolds (aSSMs) was recently used to model and control nonlinear soft robotic simulators. Here, we extend the methodology to learn aSSM-reduced models from data collected from hardware soft robotic platforms. Specifically, we learn a five dimensional (5D) aSSM-reduced nonlinear ordinary differential equation (ODE) that accurately models the dynamics of a soft-helicoid trunk robot. We then use this ODE in a model predictive control scheme to demonstrate real-time precision control across various challenging closed-loop trajectory-tracking tasks, ranging from 2D to 3D trajectories with static and dynamic constraints. We find that our methodology enhances accuracy and precision across all control tasks compared to available data-driven linear methods and model-free reinforcement learning. While we focus here on a soft-helicoid trunk robot, our methodology relies solely on hardware data and is therefore applicable to a broader class of soft robotic systems.
Understanding 3D scene dynamics in minimally invasive surgery (MIS) is a fundamental step toward physically intelligent and autonomous surgical systems. However, existing public datasets offer limited scalability and diversity in real-world clinical scenarios. Moreover, annotations typically focus on either tissue or instruments alone, overlooking tool-tissue interactions. This gap limits applications such as augmented reality guidance and autonomous surgery driven by vision-language-action (VLA) models or world models, which demand large-scale, high-quality 3D supervision. In this work, we construct SurgTAP-3D, a large-scale in-vivo pseudo-3D surgical motion tracking dataset and benchmark, comprising 520K frames, 4,549 action clips, and 1.4M 2D/3D point trajectories from diverse robotic and manual procedures under unified action categories. To make annotation scalable, we design a unified framework for surgical 3D point tracking that converts uncalibrated monocular surgical videos into structured 4D annotations through contact-region prompting, active-tool segmentation, dynamic bundle adjustment, and global 4D refinement. This pipeline adapts foundation priors to surgical videos and produces temporally consistent trajectories, pseudo-scale geometry, camera poses, and instrument masks. The dataset captures realistic challenges inherent to surgical environments, including occlusions, smoke, bleeding, and rapid instrument motion. We establish a comprehensive benchmark on SurgPose, SurgT, and STIR, covering both instrument and tissue tracking. We demonstrate instrument displacement estimation, deformable surface reconstruction, and foundation-model fine-tuning. Extensive experiments and analysis demonstrate the effectiveness of our dataset and annotation framework for 3D surgical point tracking, reconstruction, and surgical action understanding. Homepage: https://wrld.github.io/surgtap-3d/
Self-driving laboratories (SDLs) integrate laboratory robotics and automation with AI-driven decision-making to close the design-build-test-learn (DBTL) cycle. However, physical testbeds that combine off-the-shelf hardware and AI into a functioning closed-loop workflow remain in early stages of development, and their integration with commercial laboratory automation is comparatively less developed. In this work, we develop a physical testbed that assembles and integrates commercially available laboratory automation instruments and demonstrate continuous, closed-loop, autonomous operation over 24 cumulative hours of unattended experimentation. We validate this platform on a minimal closed-loop discovery task in which the system autonomously formulates aqueous dye mixtures to match a target absorbance spectrum. The system completed six independent campaigns, each consisting of 48 closed-loop iterations over approximately four hours. In every campaign, the system autonomously selected, dispensed, and measured each mixture, with an AI-driven decision-making algorithm determining subsequent formulations without human intervention. Using this testbed, we compared the performance of two AI-driven decision-making algorithms, traditional Bayesian optimization (BO) and an agentic approach, in selecting mixtures under a fixed total-volume constraint. The agentic optimizer matched target absorbance spectra using fewer iterations than BO, with reductions up to 52%, and converged on a solution for one target that BO failed to match within 48 iterations. These results demonstrate a reference physical platform for SDLs that can evaluate both established and emerging AI strategies, including large language model (LLM)-directed optimization, before their deployment in real-world applications where experiments may be costly and time-consuming to prepare.
Failure monitors for vision-language-action (VLA) policies are trained offline and then frozen. After deployment, lighting, objects and cameras drift, and the policy itself is updated or moved to new tasks. On recorded real-robot runs, after a switch of task or policy version, a frozen monitor fit on the old task ranks failures little better than chance. We present Streaming Failure Forecasting (SFF), a monitor that keeps learning from the one label every episode provides: whether it succeeded. SFF keeps the offline monitor's network fixed and refits only its last linear layer after each episode, at 2.18 ms per policy call on a Jetson CPU core. Its alarm threshold moves only after successes, which in simulation kept false alarms near target except when two drifts came at once. With every setting locked before testing, SFF ranks failures better than the same monitor kept frozen on 4 of 5 simulated test suites under unannounced drift, and above resetting on detected drift on every suite. Carried with its locked learner settings to 434 recorded runs of four VLAs on real SO-101 arms, replayed across unannounced switches of task or policy version, it ranks failures better than the frozen monitor for every VLA and, in an exploratory comparison, above a full reset at the true switch. A deployed failure monitor need not stay frozen: refitting its readout from outcomes is cheap and helps after the changes we tested.