Robotics & embodied AI

Robot learning, manipulation, navigation, vision-language-action models, embodied agents in the physical world

191 articles · page 1 of 4

Robotics & embodied AI

Language-aligned latent actions scale robot learning from human videos

The authors present SALA, a method that uses language descriptions of atomic actions to align latent action spaces of unpaired human and robot videos. This alignment allows the same policy to be trained on both sources, enabling scaling along two axes: pre-training on increasing amounts of general human video consistently reduces the robot's action prediction error on downstream tasks, and post-training on more task-specific human video substantially improves real-world robot task success rates. In experiments, pre-training on up to 2,678 hours of human video showed consistent improvements, and post-training lifted success from 53.75% to 72.5% on a 63-DoF real robot.

·3 min
Robotics & embodied AI

Robot memory that runs by clock confuses waiting with working

Zhang and An show that memory-augmented VLA policies that sample frames uniformly over time cannot tell waiting from working. A five-second pause or a wrist fidget reduces success from 47/50 to 16/50 on one counting task and from 44/50 to 10/50 on another, a phenomenon they call 'phantom progress'. The failure is traced to a temporal shortcut: because demonstrations proceed at a steady pace, the policy learns to read progress from elapsed time alone.

·2 min
Robotics & embodied AI

Robots can learn manipulation from human video via four paradigms

This review defines Ego2Robot learning as the process of adapting robot policies from egocentric human video. The authors organize methods into four paradigms: explicit embodiment transfer, explicit action supervision, representation pretraining, and latent action modeling. They find that video scale alone does not guarantee usefulness; the retained interaction information must support robot control.

·2 min
Robotics & embodied AI

Simple retrieval beats most trained robot policies on a popular benchmark

Grachev et al. show that a training-free k-nearest-neighbors (kNN) policy that simply retrieves demonstration states similar to the current observation and replays their action chunks achieves 91.7% success on the LIBERO benchmark, comparable to several trained vision-language-action (VLA) models. Under perturbations (LIBERO-Plus), the retrieval policy shows the smallest performance drop of all evaluated policies, suggesting that much of the benchmark's success may stem from memorization rather than generalization.

·3 min
Robotics & embodied AI

Inverse dynamics initializes better action sequences for visual planners

The authors introduce Dynamics-Informed Action Learning (DIAL), a framework that trains an inverse dynamics model to infer actions from transitions in a latent world model, and uses it to initialize action search for goal-conditioned planning. This bridges a gap between how latent dynamics are learned and later used for planning, leading to consistent improvements in goal-reaching performance across four visual control benchmarks, particularly with long horizons or small search budgets.

·3 min
Robotics & embodied AI

Human motion data helps humanoid robots estimate their own movement

The authors train a unified inertial encoder on human motion and kinematically retargeted humanoid motion, enabling zero-shot transfer to a physical Unitree G1 humanoid without real-robot training or fine-tuning. The model achieves the lowest velocity error among evaluated methods across five datasets, including 0.099 m/s on the real G1, and extends motion-description retrieval to robot motion categories seen only in human training data.

·3 min
Robotics & embodied AI

Humans intervene when robots stall, not when they err

The authors analyze real human intervention data in robot manipulation tasks and find that humans intervene when the robot's progress over a time horizon stalls, rather than reacting to instantaneous suboptimality. They propose a progress-based synthetic human model that better matches real behavior and introduce Q-OIL, an algorithm that adds an optimism bonus to intervention transitions. In simulation and on real robots, Q-OIL learns more efficiently, achieving 10–50% higher success rates than competing methods, and also improves pretrained vision-language-action policies on bimanual tasks.

·3 min
Robotics & embodied AI

Human corrections on rollback segments help desktop robots reach 70% success

The authors present Meituan-Robotics-0 (MR0), a vision-language-action foundation model and embodied agent for stationary-base desktop manipulation. They introduce Rollback, a data collection mechanism that pairs replaced policy segments with human corrections to provide positive and negative supervision for CFG-DAgger without value learning, achieving 70% success and a 0.90 progress score on Ethernet insertion. MR0 also combines 147 million vision-language samples with 15,060 hours of robot trajectories and uses a three-stage training recipe to align multimodal representations with robot behavior.

·3 min
Robotics & embodied AI

Multi-agent framework learns to coordinate puzzle piece assembly without planning

Knaußt et al. introduce GOMA, a multi-agent reinforcement learning framework that represents each geometric part as a cooperative agent, using a graph neural network to model spatial relationships and predict coordinated motions. In Tangram assembly tasks, the learned policy successfully completes configurations with tight clearances, outperforming a sampling-based planner and a joint-action baseline.

·2 min
Robotics & embodied AI

Simulated humanoid robots learn to walk with 81% fewer falls

Nico Bohlinger and Jan Peters show that simulated humanoids can learn to walk using a reward based on onboard acceleration measurements rather than externally measured body velocity. Their approach learns within 14 minutes of simulated experience and reduces training falls by 81% relative to a setup with weaker joint control and no left-right symmetry augmentation. Physical-robot learning remains untested.

·2 min
Robotics & embodied AI

Robot policy outperforms larger generative models with a quarter of the parameters

The authors introduce La-JEPA, a world-action model trained from scratch in the feature space of a pretrained video encoder. It predicts future visual features jointly with actions during training but only generates actions at inference. Among policies without embodied pretraining, La-JEPA achieves the best results on RoboTwin 2.0 and corrected LIBERO-Plus benchmarks, using far fewer parameters and running up to 5× faster than the strongest competing generative-backed policies. On real-robot tasks, its average success is comparable to the embodied-pretrained pi0.5.

·3 min
Robotics & embodied AI

Metric depth helps robotic manipulation policies generalize across different robots and settings

The authors trained a world action model that jointly predicts future RGB video, metric depth, and robot actions by incorporating stereo-derived depth from handheld gripper cameras. They also built a 20-task simulation benchmark to test policy robustness to changes in environment and robot embodiment. The model, Spatial-UMI-Dream, achieved 79.1% in-domain success and 72.2% under combined shifts, exceeding strong baselines by up to 10.3 points, and showed the smallest performance drop when transferring to a different robot.

·3 min
Robotics & embodied AI

Robot policies that read early vision layers handle camera shifts better

Fan et al. introduce NEO-VLA, a robot control model with no separate vision encoder: raw image patches are fed directly into a shared vision-language backbone along with the instruction, and an action module reads intermediate representations from several depths. They report that action-relevant visual signals often peak well before the final representation, and that reading those earlier layers improves control, especially when the backbone is frozen.

·2 min
Robotics & embodied AI

Robots can improve on untrained tasks by practising diverse ones

Korniak and colleagues studied whether a generalist robot policy, a model controlling many different tasks, can improve broadly rather than merely specialize during adaptation. They report that learning from interaction across diverse manipulation tasks can also improve performance on tasks excluded from that adaptation, with the strongest results coming from methods that learn to estimate the value of actions.

·2 min
Robotics & embodied AI

Soccer-playing humanoid robot learns four skills from human motion data

The authors introduce a hierarchical reinforcement learning framework that distills a latent motion prior from human soccer motion capture data, allowing a single policy to control a humanoid robot across moving, dribbling, trapping, and kicking. The system transfers zero-shot to a real Unitree G1 robot, demonstrating agile and natural execution of all four skills without separate controllers.

·2 min
Robotics & embodied AI

World model simulator works across 22 robot platforms, enables policy learning

The authors present PrimeSim, an action-conditioned multiview world model that acts as a robot simulator. Trained on over 30,000 hours of demonstrations from 22 robot platforms, it achieves faithful action following, consistent views, and long-horizon stability. PrimeSim outperforms existing world models on WorldSimProbe and EWMBench and can be used to improve robot policies via GRPO training in simulation.

·2 min
Robotics & embodied AI

Gameplay dataset maps tasks to subgoals for training AI agents

The authors present AceAction-5M, a dataset of over 5.4 million human gameplay episodes from 42 3D games, each annotated with hierarchical instructions: a task-level description, stage-level subgoals, and frame-level action instructions aligned with keyboard-and-mouse input. The key contribution is making subgoal transitions explicit with temporal boundaries, which existing datasets lack. An audit of 500 trajectories achieved a 95.8% pass rate in independent human review.

·2 min
Robotics & embodied AI

Bimanual robot learns precise tasks without any robot-collected data

Wu et al. introduce AIRoA-0, a bimanual manipulation system trained entirely from 16,151 hours of handheld demonstrations recorded with an open-source gripper. The system performs long-horizon tasks like sorting parts, folding towels, and unzipping pencil cases, using only wrist cameras and with no robot data, teleoperation, or reinforcement learning. On ten tasks at an unseen site, it achieves 78% success, compared to 27% for the best baseline.

·4 min
Robotics & embodied AI

Robots can learn hidden billiards physics by choosing their own shots

Zhang and colleagues present a robotic system that learns unfamiliar billiards dynamics by choosing interactions, observing the results and revising its simulation. Across 25 simulated conditions with hidden physical properties, it reduced held-out ball-position error by 72% relative to the strongest baseline. The authors also deployed the system on a real pool table, although the supplied text gives no comparable numerical result for that deployment.

·2 min
Robotics & embodied AI

Robots can borrow a video model's foresight and act 21 times faster

World action models improve robot control by having a large video model imagine future frames, but that imagination costs compute at exactly the moment a robot needs to react. Zhang et al. propose distilling the useful part of the video model's internal representations into an action model during training, so the video model is never run at test time. Their system matches strong success rates in simulation and real-world manipulation while running inference about 21 times faster than a representative world action model baseline.

·3 min
Robotics & embodied AI

Unlabelled navigation videos can teach robots how scenes change as they move

Navigation world models predict the next camera view given a proposed movement, letting an agent imagine a route before driving it. These models normally need trajectories where the actions were logged, which limits how varied the training data can be. Nie, Song and colleagues show that ordinary action-free video can stand in, by inferring a compact "what moved" code between frames and later aligning it with physical actions.

·3 min
Robotics & embodied AI

Robots can act at 50 Hz while updating visual understanding once a second

The authors present Leap-VLA, which decouples semantic and control loops in vision-language-action (VLA) models. A large vision-language model runs at a low rate (e.g., 1 Hz) as a semantic anchor, while a flow-matching action expert operates at control frequency (20-50 Hz) by reusing the cached semantics and consuming a fresh, low-latency visual stream. The approach recovers over 94% success rate compared to a synchronous baseline, even with semantic updates slowed to 2 seconds, demonstrating that staleness can be effectively managed.

·3 min
Robotics & embodied AI

Robots learn dexterous skills from a single egocentric human video

Gao et al. propose HandMimicX, a reinforcement learning framework that extracts a generalizable dexterous manipulation skill from a single egocentric RGB-D human video. The method focuses on object motion rather than hand trajectories, and uses interaction priors to ensure natural hand configurations. In simulation and on a real robot, the policy transfers to novel spatial configurations with improved success rates over prior object-centric approaches.

·3 min
Robotics & embodied AI

Simulator bugs cause most robot policy failures in common benchmarks

A study by Tsinghua University researchers finds that many failures in popular robot simulation benchmarks are caused by the simulator, not the policy being tested. In five container placement tasks, over 90% of a leading policy's failures involved the simulator ejecting an object the robot had placed correctly. The root cause was traced to a fixed threshold in the collision detection algorithm of PhysX 5.3.1, and changing one constant eliminated all such artifacts.

·3 min