Learning, explained and examined
How Robots Learn explains the data, methods and evidence behind robot learning, from demonstrations and simulation to real-world evaluation.
Reliability and safety
Separating pose-goal attainment from object drops revealed the reported tradeoff: the state-based LEAP-hand policy reached 95.2% of goals, while a residual safety filter cut drops by 75% and retained 91.5% goal attainment.
Written
Data and benchmarks
RobotAgentBench evaluates open-world household agents at two levels: five task-level dimensions and scenario-level integration around high-level user needs. Its real-to-simulation test set covers 50 layouts, multiple task horizons and four open-ended scenarios.
Written
Embodied AI models
ThinkWorld initializes discrete Chain-of-Dynamics motion tokens from action-labeled demonstrations. It then predicts the future induced by each chain and reinforces chains whose predictions better match observed video outcomes, allowing unlabeled human and robot videos to supervise a shared action-reasoning space.
Written
What changed
Research updates
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.
On this site by
Research updates
RLHND converts a pretrained Cosmos 3 video diffusion backbone into a deterministic clip-level feature extractor for monocular egocentric recordings. Separate streams estimate anatomically plausible hand pose and dense surface contact and force; the authors also claim leading benchmark performance and robot-learning utility, but the available abstract contains no numerical or task-level evidence.
On this site by
Research updates
Bee uses the consistency of human corrections to constrain a robot policy, with results from three physical tasks and one simulation.
On this site by
The foundations
Learning methods
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.
Published
Learning methods
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.
On this site by
Learning methods
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.
On this site by
Beyond the lab
Embodied AI models
PaceVLA uses mode-specific decoders, a quantizability signal and a context-conditioned router to choose between coarse and fine temporal action modes. Its authors report better execution efficiency without reduced task success in simulation and six physical-robot tasks, but the admitted evidence contains no detailed metrics or independent confirmation.
Published
Embodied AI models
SERA maintains task context with observations and execution feedback while a fixed VLA model handles low-level control. Environment rollouts evaluate changes to its high-level planning harness, and the selected configuration is frozen before held-out testing.
Published
Simulation to reality
PhysCanvas unifies multi-view RGB, tactile images, marker displacement, task state and actions in a shared vision space. Its frozen reward models supervise scene preservation, visual-tactile compatibility and agreement with real task outcomes. Across four real-world manipulation tasks, the authors report increased mean task-wise correlation between world-model and real-world success rates and a higher mean real-world success rate after reward-guided policy post-training.
Published
Proof and limits
Data and benchmarks
Counterfactual pairs compare similar present observations whose different histories require different actions, providing a targeted behavioral test of whether a robot policy responds to relevant past events.
Published
Data and benchmarks
A useful dexterous-manipulation benchmark separates demonstration-supported execution from held-out transfer, varies embodiments and task demands, diagnoses distinct execution failures, and keeps simulation scores separate from qualitative real-robot evidence.
Published
Data and benchmarks
LADDER holds the environment constant while scaling manipulation tasks from one to ten subgoals. Its authors report that direct low-level policies collapse on longer tasks, orchestration improves results, and remaining errors reveal limits in physical execution, though the evidence is confined to their preprint abstract.
Published