Overview

Robot policy learning studies how robots can turn camera, depth, tactile, and proprioceptive inputs into actions for tasks such as manipulation, locomotion, and navigation. Common approaches include behavior cloning from demonstrations, reinforcement learning in simulation or on hardware, and diffusion, flow-matching, or vision-language-action policies that generate action sequences; world models and hierarchical planners can add foresight and long-horizon structure. Researchers evaluate these methods in simulation suites such as LIBERO, CALVIN, RoboTwin, and navigation benchmarks, as well as on physical robots, with growing attention to memory, contact, safety, and recovery from failure. A central tension is how to make policies generalize across tasks, robots, and changing conditions while remaining data-efficient, reactive, reliable, and fast enough to deploy.

Latest papers

405 papers42 papers in the last four weeks, down 58% on the four weeks before. 0.3% of all new papers.

Nov 18, 2026

RecallFlow: Episodic Memory as the Source Distribution of Generative Robot Policies

Diffusion and flow-matching policies have become a standard choice for visuomotor imitation learning, owing to their ability to model multimodal action distributions. In memory-dependent manipulation, the robot must act on information that is no longer visible. Existing methods feed the observation history to the policy as an additional condition, while action generation still starts from Gaussian noise. However, the policy must then pick out the relevant target from a long and redundant history, leading to spurious correlations and unreliable decisions as the memory load grows. Additionally, the initial noise carries no information about what the robot has seen, leaving generation without guidance toward the remembered target. To address these issues, we propose RecallFlow, a flow-policy framework that builds the source distribution from episodic memory. Recall-Relocalization relates the memory to the current view through a few recall slots, so that the source reflects where the remembered targets are now. Recall Anchors then turns each slot into a Gaussian component of the source, enabling generation to start close to the intended action instead of from noise. The transport network then completes the action from the current observation alone, keeping inference nearly as cheap as a memoryless policy. We further build a suite of memory-dependent tasks grouped by memory load and compare all policies under a matched protocol. Extensive experiments show that RecallFlow achieves a state-of-the-art average success rate of 80.4% on the public MIKASA-Robo memory benchmark, and outperforms the strongest baseline by 20.3 points on our suite and by 18.8 points on four real-robot tasks.
Nov 17, 2026

LoRN: Low-Rank Noise Steering for Generative Robot Policies

Noise steering improves pretrained generative robot policies by learning to select their initial noise without updating model weights. We find that actions in the evaluated world-action models are more sensitive to video noise than action noise, making video noise an expressive input for steering. However, the noise-steering methods we evaluate, in their published forms, struggle to scale to the high-dimensional noise inputs of large policy classes such as world-action models. We introduce Low-Rank Noise Steering (LoRN), a reinforcement learning method that scales noise steering to world-action models. A small steering actor selects a direction through a fixed low-rank random projection, while our norm-controlled construction keeps the resulting noise near the typical scale of a Gaussian sample. This separates the actor's control dimension from the full noise dimension, enabling steering of both video and action inputs. Across three world-action models and ten model-task pairs, video steering improves average success by 26.2 percentage points over the base policy, compared with 7.1 points for DSRL action steering. For smaller action policies, the same construction achieves comparable final success while generally improving early-training success. Steering actors trained in simulation also transfer to three physical bimanual manipulation tasks across two base policies, without additional real-world training.
Nov 17, 2026

Progress Predicts Policy

A task specifies how the world should change, while a policy specifies what the robot should do. We introduce Progress Predicts Policy (PPP), which learns a language-conditioned scalar field through the commands induced by its differential. We obtain action conditioning by composing a latent progress field with a shared world model and read its action gradient at a fixed reference command. Continuous commands scale this gradient by a learned positive gain, while discrete commands use finite differences of the same scalar. This yields a hybrid policy without a separate action head or iterative search. Fitting demonstration progress values alone does not ensure that action gradients align with expert commands. PPP trains the induced commands by imitation and uses double backpropagation for the continuous gradient readout to shape the field's local control geometry. Demonstration timing calibrates field values on observed states without additional annotation. The continuous readout remains learnable without progress labels but these timing-derived targets support progress monitoring during execution. Experiments in simulated and real robot manipulation show that PPP outperforms behavioral cloning under a shared architecture.
Nov 16, 2026

ACiT: Proprioceptive-Prior Action Chunking with Inverted Transformers

Visuomotor policies condition action prediction on visual and proprioceptive observations. While historical proprioceptive sequences carry rich dynamic priors, current methods treat them as equal to high-dimensional vision, causing the proprioceptive signal to be overwhelmed. To identify what is lost in equal-conditioning fusion, we analyze proprioceptive data and find that strong inter-joint coupling is the key action-determining structure. This finding aligns with the neural sensorimotor loop, in which proprioception first forms a body-centered prior and vision is later calibrated via parietal remapping. Inspired by this, we propose ACiT, Proprioceptive-Prior Action Chunking with Inverted Transformers, which reinforces historical joint-coupling as the primary condition with vision as a later calibration. This design is instantiated through three modules: Inverted State Encoding builds a proprioceptive prior via per-joint tokens and inter-joint self-attention. State-Visual Grounding lets joint tokens query visual evidence, making vision goal- and calibration-only. Inverted Action Chunking decodes per-joint trajectories independently. ACiT improves success rates on most LIBERO, RoboTwin 2.0, and real-world tasks, with the largest gains on those that require dynamics and coordination. The same conditioning design transferred to the VLA backbone pi_0-ACiT yields significant gains with better execution efficiency.
Nov 15, 2026

RoboKit: Accelerating Parallel Differentiable Optimization for Robot Learning

We present RoboKit, a high-performance, general-purpose toolkit for parallel differentiable optimization in robotics, built on three principles: (i) Performance and flexibility: GPU-parallel multi-seed solvers pair sampling with versatile optimizers, running 2-60×\times faster than task-specific baselines across seven tasks; (ii) Geometry, rendering and physics: hybrid and coarse-to-fine representations accelerate collision, differentiable rendering and simulation, optimizing 4096 grasps in 2.97 s; (iii) Integration with learning: costs are differentiable and simulation is parallel, so RoboKit both distills optimization into policies and trains RL; distilled policies run 9-240×\times faster at comparable accuracy, and RL reaches a GPU baseline's scores up to 2.1×\times sooner on the same simulator.
Nov 14, 2026

Enforcing Ellipsoidal Constraints in Neural Networks with a Polynomial Safety Layer

Often the output of a learned mapping must satisfy prescribed constraints. In the control of autonomous systems, for example, the chosen action must keep the system within a collision-free region. We consider the case in which a proposed output is corrected by projection onto a safety ellipsoid. Computing this projection amounts to solving a quadratically constrained quadratic program (QCQP). We construct a neural-network safety layer by unrolling a feasibility-preserving polynomial penalty iteration. The unrolled structure allows the loss gradient to propagate from the feasible output through the safety layer to the network parameters during training. We show how to choose the layer parameters so that every iterate is feasible and give a computable bound on its distance to the exact projection at any prescribed depth. We illustrate the method by training a LiDAR-based policy to control an unmanned aerial vehicle. A privileged model predictive control expert uses the map and the measured vehicle state to generate the training data. In the simulated narrow mazes, the filtered policy avoided wall contact at every tested stage of training, including at initialization, when every unfiltered flight hit a wall. A representative build of the learned policy and its safety layer for an Arm Cortex-M4F microcontroller occupies about 36 KiB of flash, with a RAM budget below 5 KiB.
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

BlinkMimic: Training a Humanoid Motion Tracker in Minutes on a Single GPU

Whole-body motion tracking lets a humanoid robot follow reference motions, but training a tracking policy for a library of motions has usually taken hours of GPU time. SuperMimic showed that the policy can instead be trained by backpropagating through a world model learned from parallel simulation, and InstantTrack added Muon updates, a residual policy around the reference and an averaged deployed policy. We build on both and present BlinkMimic, a recipe that makes this training faster. On a simulated Unitree G1, BlinkMimic trains one policy for all 40 LAFAN1 clips in 3.4 minutes on a single RTX 4090, measured until three consecutive checkpoints reach a normalized return of at least 0.82 with no failed clip (mean of three seeds). We split training time into the cost of an iteration, the iterations needed to fit the motions, and the iterations needed to remove the last failures, and assign each component, inherited or new, to one of these terms. The new components are a step size that decreases once training has progressed, resets directed by failures of the deployed policy under deployment conditions, and GPU graph capture of the whole iteration. A leave-one-out ablation with three seeds per component shows which term each component affects, and further measurements of the gradient, the update rule and the failure cases explain these effects. Code and per-seed results are available at https://github.com/fuyichen0725/blinkmimic.
Nov 12, 2026

Robot Navigation and Obstacle Avoidance from Randomly Positioned External Cameras

We study robot navigation in indoor environments from fixed exocentric cameras (infrastructure-mounted, external to the robot). We show that standard Proximal Policy Optimization (PPO), which succeeds with onboard cameras, fails to learn this navigation end-to-end in our setup. This failure results from the perspective mismatch between exocentric observations and egocentric action and goal formulations, which is too complex to resolve from the reward signal alone. To address this, we propose Viewpoint-Invariant Navigation from Exocentric Cameras (VINE). VINE uses an obstacle map estimation network, trained with supervised learning, that takes exocentric images, camera pose, robot pose, and room dimensions as input and produces a viewpoint-invariant obstacle map. This map can be used either by a PPO policy or by a planner such as A*. VINE's map estimation performs better than perspective mapping approaches, handles cameras placed across a wide range of positions and orientations without view-specific tuning, requires only approximate camera pose estimates (noise with a standard deviation of 30cm per axis in x, y, and z and about 20° of rotation error on average), is robust to intrinsic variation, and extends to two-camera setups with partial coverage. Despite operating from cameras randomized in position during evaluation, VINE surpasses the navigation success rate of PPO with an onboard camera in simulated environments.
Nov 11, 2026

APG-Planner: Learning Vision-Based High-Speed Trajectory Planning via Analytic Policy Gradient

Vision-based high-speed navigation requires a trajectory planner that can reactively generate smooth, collision-free trajectories from onboard visual observations. Deep reinforcement learning offers a promising approach to vision-based trajectory planning, as it enables the optimization of sequential decision-making policies in complex environments. However, model-free policy gradient methods rely on sample-based policy gradient estimation, which yields noisy gradient estimates and inefficient training. To this end, we propose a framework for learning trajectory planning policies using the analytic policy gradient (APG) method and use it to train APG-Planner, a vision-based high-speed trajectory planner. The proposed framework leverages differential flatness to analytically reconstruct vehicle states from planned trajectories, thereby enabling differentiable trajectory-level state transitions without requiring an explicit low-level dynamics model. By iteratively performing trajectory planning and differentiable state transitions, we construct a differentiable episode rollout and analytically compute policy gradients from episode-level losses, improving both training efficiency and policy performance. After approximately 30 minutes of training on a single GPU, APG-Planner demonstrates comparable or superior performance to existing baseline methods in simulation and achieves speeds of up to 11,m/s in real-world flight experiments. We further validate the proposed framework on a skid-steer ground robot, demonstrating its applicability to differentially flat robotic systems beyond quadrotors.
Nov 10, 2026

Ego Guidance: Enhancing Robot Policy Learning with Action-Free Human Egocentric Videos

In this paper, we investigate how human egocentric videos can benefit robot policy learning without action supervision. Building on a pretrained VLM, we propose Shared Prediction on Action and Representation (SPAR), an MoT architecture in which future visual state prediction can provide independent and complementary learning signals alongside action prediction. We find that representations learned from egocentric videos improve downstream performance on out-of-distribution robot tasks, with further gains as video data scales. We also study the effects of objective choices and data curation, providing insights into how to leverage action-free videos for scalable robot learning.
Nov 8, 2026

Off-Policy RL with Online Inference Tuning: My Solution to the CoRL 2026 RoboEval Challenge

The CoRL 2026 RoboEval challenge has eight bimanual tasks, and it ranks a policy not only on whether it succeeds but on how: every stage of the task, done safely, efficiently and with both arms in coordination, scored against human demonstrations. My solution has two parts. The generic part, carried over from my LeHome challenge solution, is off-policy RL on a π0.5π_{0.5}-based policy that is also its own value function. It is trained with AWR and RECAP on advantages of the exact official score, in a loop where every rollout is at the same time a Thompson-sampling trial of about 20 inference parameters. Out of the box, it took the policy to 0.85 RDS on held-out seeds in two days (and it was still very far from convergence). The competition-specific part adds data that local exploration cannot find: generated demonstrations that reach the full reward, and recovery episodes that start from the policy's own failures. The submitted policy reaches 0.990 RDS on held-out seeds and 0.987 on the official seeds.
Nov 7, 2026

Deploying Physically and Epistemically Compliant Agents by Distilling Belief-Space Task and Motion Planners

In this work, we train agents to act under uncertainty according to specified, belief-dependent rules. Deep reinforcement learning (RL) can in theory train agents to follow rules if encoded into well-designed rewards, but in practice, RL fails to perfectly optimize these rewards, leading to ineffective or noncompliant agent behavior. Belief-space planners by construction produce behavior in line with developer specification, but require hand-programming online planners and belief-state estimators, adding significant engineering complexity at scale. We combine the two: we plan offline to generate a dataset of sample trajectories and then distill the plans into a policy that acts on raw observations. We propose two changes to planner distillation that address problems specific to history-based policies: Dual Averaging, a loss that combines per-timestep predictive loss and average loss on task changes, and ρρ-Optimal Reinforcement, a feedback mechanism that reinforces the policy's own successful rollouts when these are near-optimal. We provide a sufficient condition under which static search problems with noiseless observation models is complete, and show that maximum-likelihood-observation planning implies this condition in static search with continuous beliefs. We create a challenging new domain, Scavenger, where an agent needs to gain information in order to succeed. We evaluate our distilled policy in two versions of Scavenger, one where information is provided visually and one where it is provided as a command in natural language. In either case, the learned policy attains a success rate above 90% in Scavenger, compared with at most 5% for RL. We also find that distilled planners transfer with greater than 60% success to following unseen and intentionally vague commands and greater than 75% success to following floorplans larger than those seen in training.
Nov 6, 2026

A Controlled Study of Models, Skills, Feedback, and Curriculum in Agentic Robotics

Agentic robotics asks coding agents to implement and iteratively improve robot policies from simulated and real-world physical interactions. We study the self-learning loop of agentic robotics through a controlled simulation study spanning nine manipulation tasks, 631 different experiment configurations, and 28K coding agent rollouts, systematically varying model capability, skill priors, feedback structure, and curriculum. The experiment data suggests that the bottleneck is often not the access to sufficient information, but interpreting it into the right policy update. Across these factors, a successful self-learning loop depends on how effectively agents attribute failures, reason task-relevant spatial and geometric structure, and selectively restructure code to improve policies.
Nov 5, 2026

CS-Pi: Constraint-Satisfying Flow Policy Sampling

A flow-matching policy deployed in the field must often satisfy a constraint that was absent from its demonstrations, such as avoiding an obstacle it has never seen. Retraining for every new constraint is impractical, so the constraint must be enforced at inference time. Enforcing it must not discard what the policy learned: an action the policy would rarely sample may avoid the obstacle yet fail the task. We therefore want actions that are both safe, satisfying the constraint, and plausible, remaining likely under the policy. Gradient guidance, the most common approach, steers noisy states, so it must evaluate its cost on the policy's running estimate of the clean action; when the policy is uncertain between passing left or right of an obstacle, this estimate is their average and can pass straight through it. We instead build on rejection and projection. Rejection redraws actions until one is safe: every output is plausible, but the wait is intractable when safe actions are rare. Projection moves an action to the nearest safe one: it is efficient, but the result need not be plausible. Our key insight is that flow-matching sampling lets us combine the two, keeping projection's efficiency and rejection's plausibility. CS-Pi, our draw-project-renoise sampler, projects each draw to the nearest safe action, adds noise back scaled to how far the projection moved it, and re-runs the unmodified policy until a draw is safe. Replacing only the projection step extends CS-Pi to continuous costs. Across a 2D diagnostic, a 7-DoF shelf pick-and-place, an orientation preference, and a zero-shot Franka deployment, CS-Pi reaches 81% safe success on the shelf and 55% on hardware, against 55% and 40% for projection, while staying plausible.
Nov 5, 2026

La-JEPA: Compact Robot Policies with Latent World Modeling

World-action models adapt pretrained video generators, and vision-language-action models build on pretrained vision-language backbones; both tie policy architecture and inference cost to large transformers designed for generation rather than control. It remains unclear how much these generation-oriented choices affect control. In this work, we introduce Latent Action JEPA (La-JEPA), a compact world-action model whose policy transformer is trained from scratch in the feature space of a pretrained V-JEPA 2.1 encoder. Following Fast-WAM, La-JEPA learns to predict future features jointly with actions during training and generates only actions at inference. Among policies without embodied pretraining, La-JEPA achieves state-of-the-art performance in clean and randomized RoboTwin 2.0 scenes and on LIBERO-Plus with correct instructions, while using far fewer parameters than the strongest competing policies. It runs substantially faster than Fast-WAM at comparable performance and outperforms the closest JEPA-based methods, LeapBot-WA and JEPA-WAM, on LIBERO-Plus and RoboTwin 2.0 without their pretrained policy backbones. On four real-robot tasks, its average success is comparable to pi0.5 and above non-embodied-pretrained baselines. Through a design study of the choices behind La-JEPA, we show that JEPA features outperform video VAE latents, most clearly out of distribution, and that training-time future prediction choices significantly impact the policy even when future prediction is skipped at inference. We correct the LIBERO-Plus instructions, as the officially released evaluator corrupts them for most variants, and re-evaluate several strong open baselines. Without a generative pretrained policy backbone, La-JEPA is better and faster than far larger policies, suggesting that robot foundation models may benefit from architectures designed for control rather than inherited from generation.
Nov 5, 2026

LMGN: A Unified Policy Learning Model for Last-Mile Grasp-Navigation

Legged robots equipped with manipulators have shown significant potential in the fields of daily life and industry in complex environments. However, existing methods mainly focus on global navigation or manipulation independently, while overlooking the last-mile navigation problem, where the robot must jointly determine the base placement, approach direction, and grasp execution to achieve reliable object grasping, and different manipulation stages require distinct control behaviors. To address these challenges, we propose Last-Mile Grasp-Navigation (LMGN), a unified policy learning model for Quadruped robot loco-manipulation. An Explicit Distance and Directional Reachability (ED-DR) representation is first introduced to explicitly represent the features of grasp-compatible base placement and approach direction. We then develop a semantically aligned Hierarchical Mixture-of-Experts (HMoE) policy that models stage-aware behaviors across the stages of base adjustment, pre-grasp alignment and grasp execution, and object lift. Finally, a new optimization method that jointly aligns navigation and manipulation, stage-dependent expert routing, and action-level consistency is proposed to optimize the HMoE policy. Extensive experiments on the Habitat and Isaac Gym simulators demonstrate that LMGN consistently outperforms state-of-the-art methods, denoting that integrating grasp-aware reachability with stage-aware policy learning significantly improves the success rates of last-mile navigation and target grasp for robots.
Nov 5, 2026

Efficient Diffusion Policy Fine Tuning with Latent Noise Representation Bridging

Pre-trained diffusion and flow policies have emerged as powerful backbones for robotics, yet adapting them to novel data via reinforcement learning (RL) remains computationally expensive, sample inefficient and lacks theoretical guarantees. Prior approaches that fine-tune the initial noise bypass costly backpropagation through time but struggle with the "noise aliasing" problem: there does not exista natural link between the noise space value and the action space value. This paper resolves these challenges by introducing the Latent Noise Linear Markov Decision Process (LNL-MDP). We derive the Bridge Equation for LNL-MDPs which eliminates the noise aliasing problem by aligning values with a shared representation between the noise and the action space. Furthermore, our LNL-MDP fine tuning framework falls under the FineTuneRL theory with near-optimal guarantee. Leveraging these insights, we propose Latent Representation Bridging for Diffusion (LARBRID), an algorithm that efficiently learns the representation to optimize the initial distribution that steers the pre-trained model to act optimally in the environment. It yields up to 2-3×\times performance increases on challenging control problems across a diverse suite of robotic locomotion and manipulation tasks. Code can be found at: github.com/Maincode-Research/LARBRID.git.
Nov 4, 2026

Reinforcement Learning Enables Scaling Generalist Robot Policy Improvement

The introduction of generalist robot policies has enabled a single, large model to learn to perform a variety of tasks from a vast pre-training corpus. While these models often exhibit powerful zero-shot generalization capabilities, they typically require an adaptation phase to achieve sufficient performance in deployment. Iterated self-improvement via reinforcement learning (RL) has demonstrated promise in enabling this adaptation, yet such approaches have been largely explored in single-task settings - where the goal is to simply improve on a narrow deployment task - which lies in stark contrast to the multi-task capabilities of the base policy. In this work we ask whether generalist robot policies are capable of multi-task, generalist improvement in deployment - targeting many diverse tasks simultaneously - and what algorithmic recipe enables this. To answer this question, we first develop a set of structured evaluations to test generalist improvement capabilities, and systematically evaluate algorithmic and practical design choices regarding value estimation, policy extraction and resource allocation. We show that not only can generalist policies be improved online across diverse tasks, but that the right recipe can enable generalization and increase performance on tasks not considered during online training. We characterize when this generalization occurs and show that both the training task distribution and the use of value-based RL are critical. To the best of our knowledge, this is among the first evidence for online generalist improvement for robot foundation models.
Nov 4, 2026

RACE: Calibrated Cumulative Scoring for Rollout-Free Robot Policy Checkpoint Selection

Assessing the deployment performance of learned robot policies during training is challenging, as directly measuring task success requires costly real-world trials for each candidate checkpoint. Averaged validation loss offers a rollout-free alternative, but its checkpoint rankings can be inconsistent with deployment performance. We analyze two limitations of this criterion: it penalizes action deviations without considering whether the policy can recover from them, and gives later reference predictions equal weight even when earlier deviations may lead execution away from the reference. We propose RACE, a novel rollout-free metric for ranking checkpoints using a fixed set of held-out successful trajectories. RACE calibrates action deviations against state-conditional thresholds derived from action differences at similar states in these trajectories. We aggregate these scores through a running product, so that a substantial deviation reduces the contribution of subsequent reference predictions. Averaging the resulting cumulative scores over time and across trajectories yields the checkpoint score. In simulation, RACE outperforms baselines across validation sources, while several baselines degrade sharply when the validation set changes. On real robots, its rankings align with deployment success across the evaluated policy backbones and tasks of varying difficulty. Perturbation analyses further show that RACE distinguishes deviations with markedly different effects on task success, even when baselines respond similarly.
Nov 2, 2026

SNAP: Semantic Navigation Among Pedestrians

Navigating crowded urban environments requires robots to avoid pedestrians while adhering to social norms such as remaining on sidewalks. Existing approaches typically address either crowd avoidance or semantic path planning in isolation. We propose a reinforcement learning (RL) navigation policy that jointly reasons about semantic terrain and dynamic obstacles from a single semantically annotated bird's-eye-view (BEV) map, without relying on explicit human-state estimation. Our policy outperforms established crowd-navigation and semantic-aware baselines and transfers zero-shot to a legged robot, demonstrating socially compliant navigation in highly dynamic real-world scenarios.
Nov 2, 2026

MAPO: Mixture of Experts with Alternating Policy Optimization for Universal Humanoid Motion Tracking

Universal humanoid motion tracking is essential for enabling humanoid robots to perform diverse tasks in human environments. To capture the wide diversity of human motions, mixture-of-experts (MoE) policies have been widely adopted, leveraging dynamic routing to enhance policy expressiveness. However, we identify that divergence between old and updated policies can be misestimated during policy-gradient updates of MoE policies, due to their dynamic routing. Because most policy-gradient algorithms regulate policy updates based on the divergence, this misestimation can lead to either overly conservative or excessively large updates, which can yield a suboptimal policy. In this paper, we propose Mixture of Experts with Alternating Policy Optimization (MAPO), which alternately optimizes the router and the experts to avoid the divergence misestimation. MAPO consistently tracks 182 motions that PPO with MoE fails to track, whereas PPO with MoE tracks only one motion that MAPO fails to track. MAPO also attains the highest tracking success rate among the compared methods and the lowest tracking error across most error metrics, demonstrating its superior performance. Videos can be found on the project website: https://mapo-corl2026.github.io/MAPO/
Nov 1, 2026

ODE Policy: Robot Learning via Action Dynamics, Not Actions

Most robot policies, whether trained by imitation learning (IL) or reinforcement learning (RL), map observations to action values at discrete control steps and leave the continuity of physical motion to be learned implicitly. We propose ODE policy, which learns the dynamics of the action rather than the action itself. A neural ordinary differential equation (ODE) learns the vector field of the action by matching the integrated actions to the targets over short action chunks in IL and single control steps in RL. A lightweight diversity head captures multi-modal behavior by randomizing the direction of the field. At deployment, each action is integrated from the previous one, so the executed trajectory is continuous by construction. We evaluate it as a visuomotor policy, as the action head of vision-language-action (VLA) and world action models (WAMs), and as the actor of model-based RL against 14 baselines on unimanual, bimanual, dexterous, locomotion, and loco-manipulation tasks. In IL, against flow matching, it reaches 5.9×\times the success, 5.4×\times the generalization, and 3.9×\times the mode balance after one epoch, and generates actions 10.8×\times faster. It mostly moves more smoothly than human demonstrations. On 14 real-world tasks across single-arm, dual-arm, dexterous, and humanoid robots, it achieves the highest success rate. In RL, it further speeds up SuperMimic for humanoid motion tracking and smooths its actions. Efficient in learning, data, parameters, and execution, ODE policy offers a promising path to scaling robot learning.
Oct 30, 2026

Nexus-VLA: Bridging Human and Robot Experiences via a Shared Hierarchical Semantic Latent Space

Large-scale egocentric human videos provide rich and scalable manipulation experience for robot policy learning, yet substantial embodiment gaps in visual appearance, kinematic structure, and action spaces hinder effective knowledge transfer. Existing methods often decouple cross-embodiment knowledge alignment from robot policy learning, making it difficult for human manipulation knowledge to directly and fully contribute to robot action learning. To address this issue, we propose Nexus-VLA, a hybrid policy learning framework that enables cross-embodiment knowledge transfer through end-to-end human-robot joint training, allowing human manipulation experience to directly contribute to robot policy learning. We first introduce a hierarchical cross-embodiment semantic alignment paradigm that progressively establishes shared manipulation semantics from task intent to fine-grained interactions, enabling more complete use of the manipulation knowledge contained in human videos. Building on this hierarchy, we design a shared manipulation latent representation that jointly encodes multi-level shared knowledge into an internal representation directly used for continuous action generation, improving policy learning efficiency while effectively leveraging human knowledge. Across RoboCasa-GR1, OOD settings, and real-world robot experiments, Nexus-VLA achieves an 80.5% task success rate on RoboCasa-GR1 and demonstrates strong advantages in human-data utilization, learning efficiency, and cross-scenario generalization.
Oct 30, 2026

∇\nablaMimic: Ultra-Efficient Humanoid Control with Differentiable Simulation

Robot policy learning is dominated by zeroth-order methods such as PPO, which treat the simulator as a black box and pay for it with enormous sample counts. First-order policy learning differentiates through the dynamics, but on contact-rich whole-body control it has rarely worked: its gradients explode. A recent world-model method showed that derivative information can make humanoid motion tracking far more efficient, using surrogate gradients of a learned model. We show that exact gradients of the simulator itself do better, once two obstacles are removed. Exact gradients contain rare unbounded samples, because contact dynamics is not differentiable everywhere and its derivative diverges near some events; and they are expensive, because the simulator's adjoint has to replay the contact solve. ∇\nablaMimic bounds these samples with median-relative per-sample gradient clipping at every step of the backward pass, and reuses each costly simulated window by multistep regression onto gradient-derived action targets, which bounds every action target independently of the gradient scale. NablaEngine, a differentiable GPU simulation engine for the Unitree G1, removes the cost: it computes MuJoCo's full contact dynamics with fused per-world kernels and the adjoint of a whole control step as a single megakernel, with analytic gradients that match float64 finite differences, 11.5×11.5\times faster than the available MuJoCo Warp adjoint. On whole-body tracking of all 40 LAFAN1 clips with one policy on one GPU, ∇\nablaMimic reaches the score reported by the state-of-the-art world-model method, 0.82, in 20 minutes instead of 1.2 h, and 0.874 in 1.8 h. SHAC, the standard first-order baseline, stops on exploding gradients within minutes. We release NablaEngine and ∇\nablaMimic as open-source infrastructure for first-order humanoid policy learning.
Oct 30, 2026

TRACE: Target-State Risk-Aligned Conditional Execution for Embodied Tracking

Embodied visual tracking requires a robot to follow a language-specified person under occlusion, look-alike distractors, and ambiguous instructions. We show that improving tracking persistence can simultaneously increase safety violations. Using a compact 0.6B vision-language-action (VLA) tracker with a frozen backbone, we update only the planner head with proximal policy optimization (PPO). PPO improves success and following rates across all three EVT-Bench settings, but increases the collision rate on distracted tracking because episodes that previously ended in target loss can instead end in contact. We introduce TRACE, an execution-time target-state risk gate that preserves the learned policy and intervenes selectively. A lightweight risk head processes 84 cues, including candidate boxes, identity margins, a learned range-and-bearing probe, and the policy's commands. When the calibrated risk exceeds a threshold, TRACE rescales the translational component of the next waypoint while leaving yaw unchanged. The executed action is identical to PPO's on more than 98% of steps. In paired evaluations on identical episodes, TRACE reduces the distracted-tracking collision rate by 3.96 percentage points, while further improving success and following rates. Matched slowdown controls without target-state timing do not reproduce this improvement. Concurrent evaluation reproduces a collision reduction of 3.77 points, compared with 0.29 points for identical-arm null pairs, indicating that the effect exceeds evaluation variability. These results show that target-state-aware execution can recover the safety cost of policy post-training with minimal intervention.
Oct 29, 2026

Gaussian Distillation of Flow Policies

Flow policies are among the strongest robot controllers and are well suited to modelling multimodal action distributions. Our analysis shows that the continuous-action distributions of the studied flow policies are well approximated by Gaussians. We propose the Gaussian Action Readout (GAR), which replaces a flow policy's final layer with a Gaussian readout trained on just 0.02% of eligible states, while keeping the rest of the network frozen. It samples with one action-network pass, gives closed-form likelihoods, and achieves near-teacher task success at up to twice the inference speed.
Oct 29, 2026

Learning Safe Flow Matching Robot Policies from Binary Feasibility Oracles

Robot foundation models increasingly rely on flow-based action generation for general-purpose manipulation, yet the generated actions are not guaranteed to satisfy physical constraints such as joint limits and collision avoidance. This problem is especially challenging when the feasible set cannot be described analytically and is accessible only through black-box feasibility oracles. We study this practical oracle-only setting and propose an oracle-guided differentiable safeguard for safe flow-matching robot policies. The safeguard uses binary feasibility queries to correct unsafe updates in the forward pass, while extracting local boundary information to provide constraint-aware learning signals in the backward pass. Combined with discrete-time flow matching, the same safeguard is used during training and inference, enabling forward--backward aligned policy learning without explicit or differentiable constraint models. Experiments on obstacle-aware robotic manipulation demonstrate 100% constraint satisfaction while preserving task performance.
Oct 28, 2026

Learning Event-based Obstacle Avoidance via Differentiable Simulation

Reliable obstacle avoidance is essential for safe quadrotor flight in unknown environments, yet conventional frame-based cameras remain challenged by motion blur, limited dynamic range, and latency under fast motion. Event cameras, with sub-millisecond latency, high dynamic range, and negligible motion blur, offer a promising alternative, but leveraging them for closed-loop control remains difficult. Existing approaches either adopt cascaded architectures that first estimate depth and then map it to control commands, introducing error propagation, or achieve end-to-end control through costly offline pretraining before online fine-tuning can even begin. We address these challenges with a CUDA-based event simulator that exploits GPU parallelism to render events efficiently within the training loop, integrated into a differentiable simulation pipeline that enables direct gradient backpropagation through the differentiable physics model, improving sample efficiency. This allows us to train an end-to-end policy that maps event frames directly to control commands, without intermediate depth estimation. We validate our approach through zero-shot real-world deployment on a quadrotor, achieving robust obstacle avoidance in both odometry-assisted and odometry-free settings, including under low-light conditions where conventional cameras typically fail.
Oct 28, 2026

CSO: Refining Robotic Policies via Skill Distribution Alignment and Skill-Grained Optimization

Discretizing continuous actions into discrete skills using methods like VQ-VAE has emerged as a powerful paradigm for robotic manipulation. However, quantization errors yield a suboptimal distribution for the skill prior, degrading its performance. While reinforcement learning offers a path for refinement, its direct application faces unstable encoder updates and a granularity dilemma in importance sampling. To address these challenges, we introduce Cascaded Skills Optimization (CSO), a two-stage post-training framework. First, Rejection-Sampling Supervised Fine-tuning aligns the model's observation-to-skill mapping with the distribution of successful online trajectories via supervised fine-tuning. Second, with the observation encoder frozen, Skills Policy Optimization computes an independent, clipped importance ratio for each complete skill, enabling stable and efficient updates. With a lightweight QueST-style policy, CSO raises overall LIBERO success from 80.1% to 97.5%, improves performance on RoboTwin2.0, and achieves 61.0% success across five physical single-arm tasks. Additional comparisons with FAST, discrete BEAST, and fixed-buffer IQL further validate the effectiveness of our two-stage post-training strategy.