https://howrobotslearn.comHow Robots Learn explains the data, methods and evidence behind robot learning, from demonstrations and simulation to real-world evaluation.en-UShttps://howrobotslearn.com/en/articles/detecting-task-priors-in-streaming-robot-evaluation/https://howrobotslearn.com/en/articles/detecting-task-priors-in-streaming-robot-evaluation/A benchmark can distinguish temporal task understanding from prior-driven guessing by scoring recognition, matching, ordering, prediction, correspondence, and localization separately, then repeating matched questions without visual observations. The reported matching and ordering scores diverge sharply, and next-action prediction changed far less than current-action recognition when observations were removed.https://howrobotslearn.com/en/articles/separating-spatial-grounding-from-manipulation-failures/https://howrobotslearn.com/en/articles/separating-spatial-grounding-from-manipulation-failures/Spa-Bench compares in-distribution and out-of-distribution spatial configurations while retaining training support for relevant objects, concepts, and actions. The authors use diagnostic conditions to probe physical execution, sensitivity to wording, object-task associations, and recognition that no action is required; matched direct instructions improved success in the reported results.https://howrobotslearn.com/en/articles/how-demonstrated-object-effects-become-editable-robot-programs/https://howrobotslearn.com/en/articles/how-demonstrated-object-effects-become-editable-robot-programs/A demonstration is compiled into coordinates describing object motion, contact locations and the desired final scene. It holds that effect program fixed while generating an execution from the live scene, so people can edit the intended outcome and the robot can re-solve the task after changes.Sat, 17 Oct 2026 12:53:56 GMThttps://howrobotslearn.com/en/articles/what-makes-real-robot-evaluation-reproducible-and-auditable/https://howrobotslearn.com/en/articles/what-makes-real-robot-evaluation-reproducible-and-auditable/The admitted excerpts support coordinating data collection, scene reset, policy execution, result logging, and task-factor analysis, while using success classifiers and reset policies as mechanisms for repeatable trials; they do not provide a complete implementation recipe.Sat, 17 Oct 2026 12:53:56 GMThttps://howrobotslearn.com/en/articles/how-whole-body-touch-adapts-humanoid-vla-policies/https://howrobotslearn.com/en/articles/how-whole-body-touch-adapts-humanoid-vla-policies/Uni-VLaT adds a distributed tactile pathway to pretrained VLA policies and trains it to predict future touch, proprioceptive, and visual representations alongside actions. Its authors report a 75% average success rate across five real-robot tasks, while the single-preprint evidence does not provide independent confirmation.Sat, 17 Oct 2026 12:53:56 GMThttps://howrobotslearn.com/en/articles/how-supervised-play-helps-robots-learn-dexterous-manipulation/https://howrobotslearn.com/en/articles/how-supervised-play-helps-robots-learn-dexterous-manipulation/DSP behavior-clones a robot play policy from minutes of human-guided activity, deploys it to collect hours of interaction data, and trains a world model for planning and self-play policy distillation. Its authors report physical-robot results on pushing and in-hand reorientation, including a 69.6% success-rate advantage over hindsight behavior cloning and comparable success to reinforcement learning with 490 times less interaction data.Fri, 16 Oct 2026 13:12:37 GMThttps://howrobotslearn.com/en/articles/how-one-robot-policy-combines-pushing-tipping-and-grasping/https://howrobotslearn.com/en/articles/how-one-robot-policy-combines-pushing-tipping-and-grasping/Uni-PNP learns a unified object-pose-reaching policy by modeling contact-induced motion and progressing from graspable to ungraspable objects during reinforcement learning. Its authors report gains on seen and unseen objects, while providing only an unspecified claim of real-world transfer in the available abstract.Fri, 16 Oct 2026 13:12:37 GMThttps://howrobotslearn.com/en/articles/raw-replay-data-versus-lerobot-demonstration-dataset/https://howrobotslearn.com/en/articles/raw-replay-data-versus-lerobot-demonstration-dataset/Raw HDF5 files retain the information required for exact trajectory replay in OmniGibson and for collecting additional visual observations. The corresponding LeRobot release presents 20,000 demonstrations across 100 tasks in a customized LeRobot V3 structure with documented RGB and depth data.https://howrobotslearn.com/en/articles/how-residual-memory-helps-robot-policies-retain-history/https://howrobotslearn.com/en/articles/how-residual-memory-helps-robot-policies-retain-history/TTT-RM trains an online fast-weight state on the part of task history that cannot be reconstructed from a robot’s current observation and bounded memory bank. Its authors report simulation gains and a three-minute, eight-stage physical stowing demonstration, but the evidence provided is limited to their preprint abstract.https://howrobotslearn.com/en/articles/how-rgb-and-depth-are-stored-in-robot-demonstration-data/https://howrobotslearn.com/en/articles/how-rgb-and-depth-are-stored-in-robot-demonstration-data/In the LeRobot demo dataset, RGB observations have shape `(height, width, 4)`, use `numpy.uint8`, range from 0 to 255, and have resolutions of 720 × 720 for the head camera and 480 × 480 for wrist cameras. Linear depth measures the distance from the camera to scene geometry, has shape `(height, width)`, uses `numpy.float32`, ranges from 0 to 10 meters, and is stored in videos that are log-quantized during replay and then dequantized to metric depth values.https://howrobotslearn.com/en/articles/how-robotic-hands-use-fingers-for-locomotion-and-manipulation/https://howrobotslearn.com/en/articles/how-robotic-hands-use-fingers-for-locomotion-and-manipulation/A reinforcement-learning approach coordinates an anthropomorphic hand’s unequal fingers for self-support and locomotion without changing their design. The authors separately report simulated speed gains and physical demonstrations of untethered movement, recovery, keyboard input and visually guided object pushing.https://howrobotslearn.com/en/articles/when-more-human-demonstration-data-improves-robot-learning/https://howrobotslearn.com/en/articles/when-more-human-demonstration-data-improves-robot-learning/The EgoVerse authors report that robot policy performance generally improves with more human demonstrations when the data align with the learning objectives, although the available abstract does not quantify the gains or define alignment.https://howrobotslearn.com/en/articles/why-imitation-learning-errors-accumulate/https://howrobotslearn.com/en/articles/why-imitation-learning-errors-accumulate/Errors made by an imitation policy can accumulate until the robot drifts beyond the distribution covered by its training demonstrations, a covariate-shift failure mode linked to poor generalization. One real-robot chopstick study addresses single-step errors with added support at critical grasping stages and error accumulation with synthetic corrective labels, while a later preprint evaluates a policy-calibrated noise method in simulated environments.https://howrobotslearn.com/en/articles/how-robot-dataset-diversity-affects-generalization/https://howrobotslearn.com/en/articles/how-robot-dataset-diversity-affects-generalization/Variation across tasks, environments, arms, and grippers can widen the conditions a robot policy handles, but the available studies do not isolate diversity from dataset scale or establish universal gains.https://howrobotslearn.com/en/articles/how-robot-orchestrators-and-policies-improve-together/https://howrobotslearn.com/en/articles/how-robot-orchestrators-and-policies-improve-together/Robo-COP proposes a shared improvement cycle in which deployment executions provide demonstrations, recurring failures guide policy fine-tuning, and revised skills must pass verification before the new policy is adopted. Its authors report gains on ten simulated tasks and three physical tasks.https://howrobotslearn.com/en/articles/how-diffusion-policies-preserve-distinct-action-modes/https://howrobotslearn.com/en/articles/how-diffusion-policies-preserve-distinct-action-modes/Independent action-noise pairing can mix diffusion paths and produce averaged denoising responses that suppress distinct demonstrated behaviors. A label-free training add-on instead assigns noise to actions to preserve relatively separate routes without changing architecture or inference, although the reported results come from one preprint.https://howrobotslearn.com/en/articles/how-stage-level-metrics-reveal-bimanual-robot-failures/https://howrobotslearn.com/en/articles/how-stage-level-metrics-reveal-bimanual-robot-failures/Stage-level evaluation can show when and how a bimanual policy breaks down, separating alignment, slipping, coordination, arm-use asymmetry, efficiency and safety problems that binary success can conceal.https://howrobotslearn.com/en/articles/how-spatial-tokens-connect-robot-actions-and-future-states/https://howrobotslearn.com/en/articles/how-spatial-tokens-connect-robot-actions-and-future-states/A preprint proposes encoding scene contours, goals, action targets and future states as spatial sequences, then training one autoregressive Transformer for both action generation and action-conditioned forecasting. Its simulation and real-robot results are author-reported.https://howrobotslearn.com/en/articles/how-world-action-models-represent-motion-and-scene-change-in-3d/https://howrobotslearn.com/en/articles/how-world-action-models-represent-motion-and-scene-change-in-3d/Both occupy a shared metric frame, allowing one model interface to produce executable actions and action-conditioned forecasts of physical change.https://howrobotslearn.com/en/articles/what-is-inside-a-robot-demonstration-dataset/https://howrobotslearn.com/en/articles/what-is-inside-a-robot-demonstration-dataset/A robot demonstration dataset can align images and other observations with recorded actions, organize them into trajectories, and supply schemas, task or environment context, statistics and subset labels. These files are inputs to learning algorithms, not learned policies themselves, and the presence of rewards or termination signals alone does not independently confirm successful task completion.https://howrobotslearn.com/en/articles/how-robot-training-loaders-build-sequences-from-episodes/https://howrobotslearn.com/en/articles/how-robot-training-loaders-build-sequences-from-episodes/A loader can complete a requested window near an episode’s end by repeating the final frame as padding. That repetition differs from both newly recorded steps and the earlier context supplied through frame stacking.https://howrobotslearn.com/en/articles/how-success-guided-sampling-focuses-robot-reinforcement-learning/https://howrobotslearn.com/en/articles/how-success-guided-sampling-focuses-robot-reinforcement-learning/Success Guided Sampling prioritizes simulated task configurations around the boundary of a robot policy’s current abilities. The abstract reports that the method solved locomotion and assembly tasks that prior methods failed to solve and that distilled RGB-based policies transferred zero-shot to assembly tasks on real hardware.https://howrobotslearn.com/en/articles/how-contact-structure-transfers-demonstrations-to-robot-hands/https://howrobotslearn.com/en/articles/how-contact-structure-transfers-demonstrations-to-robot-hands/The method preserves where and when fingers contact an object instead of copying human joint motion directly. Its authors combine simulated physics refinement, morphology-independent contact retargeting and residual policy learning, then report deployment on four physical bimanual tasks without real-robot training data.https://howrobotslearn.com/en/articles/how-latent-future-states-guide-robot-actions/https://howrobotslearn.com/en/articles/how-latent-future-states-guide-robot-actions/PLaW-VLA predicts task-relevant futures in a pretrained latent representation and conditions action generation on them through structured causal attention. Its authors report benchmark gains and lower latency than generative world-action modeling.https://howrobotslearn.com/en/articles/how-synthetic-data-trains-mobile-manipulation-vlas-for-real-world-deployment/https://howrobotslearn.com/en/articles/how-synthetic-data-trains-mobile-manipulation-vlas-for-real-world-deployment/SimVLA uses synthetic action data and simulator-derived visual-language supervision for pre-training, then adds simulated policy rollouts during post-training. Its authors report zero-shot transfer to real mobile-manipulation tasks, but the available abstract does not provide detailed results or independent validation.https://howrobotslearn.com/en/articles/how-virtual-viewpoints-help-robots-perform-precise-manipulation/https://howrobotslearn.com/en/articles/how-virtual-viewpoints-help-robots-perform-precise-manipulation/SpatialHarness synchronizes a simulated scene with real robot execution and renders complementary perspectives for a frozen multimodal policy. Its authors report gains on four physical manipulation tasks, including plug insertion and Tower of Hanoi.https://howrobotslearn.com/en/articles/how-benchmarks-test-dual-arm-skill-recomposition/https://howrobotslearn.com/en/articles/how-benchmarks-test-dual-arm-skill-recomposition/A dual-arm benchmark can probe recomposition by testing unseen combinations of familiar per-arm skills, enforcing order and timing requirements, and comparing policies under shared data and evaluation conditions.https://howrobotslearn.com/en/articles/when-humans-demonstrate-correct-and-evaluate-robot-learning/https://howrobotslearn.com/en/articles/when-humans-demonstrate-correct-and-evaluate-robot-learning/A reading guide to human input during data collection, online training and evaluation, illustrated with LeRobot and Bee.https://howrobotslearn.com/en/articles/robot-planner-versus-action-policy/https://howrobotslearn.com/en/articles/robot-planner-versus-action-policy/Compare a skill selector with an action predictor, using the interfaces described by the SayCan and OpenVLA authors.https://howrobotslearn.com/en/articles/how-latent-robot-actions-transfer-across-starting-states/https://howrobotslearn.com/en/articles/how-latent-robot-actions-transfer-across-starting-states/TERRA learns latent actions from compact representations of temporal visual change. Its EAT mechanism aligns the effects produced by the same code across different initial states, while the reported evaluations remain claims from the preprint’s authors.