Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    How robots learn: demonstrations, trial and error, and data

    Most new robot skills are now learned rather than hand-programmed. Robots copy human demonstrations collected by teleoperation (imitation learning)[1] or improve by trial and error in simulation (reinforcement learning)[2]. The limiting factor is data: researchers have pooled about a million real robot trajectories,[3] far less than the text behind language models.[4]

    Editor reviewedUpdated Robotics and embodied AIArtificial intelligence

    Robots used to be programmed step by step for one job in one fixed cell. The modern approach trains a neural network “policy” that maps what the robot senses to what it does. This page explains the two main ways to train such a policy and why collecting data is still the hardest part.[1][2]

    Learning by copying

    In imitation learning, a person shows the robot what to do, often by steering it remotely (“teleoperation”), and the robot learns to copy. In a 2023 study a cheap two-armed setup learned six fiddly tasks, such as opening a translucent condiment cup, with 80-90% success after about ten minutes of demonstrations.[1]

    A known weakness is that small mistakes add up. Once the robot drifts away from anything it saw in the demonstrations, it may not know how to recover.[5]

    Behaviour cloning fits a policy to state-action pairs from expert demonstrations. Its classic failure mode is compounding error under distribution shift. The ACT method (Action Chunking with Transformers) reduces this by predicting chunks of future actions with a generative model.[5] Diffusion Policy treats the policy as a conditional denoising diffusion process. That handles multimodal action distributions, and the paper reported a 46.9% average improvement across 12 tasks from 4 benchmarks.[6] Flow matching, a related generative method, is the action head in π0.[7]

    Learning by trial and error

    In reinforcement learning (RL), the robot tries actions and gets a reward signal when it does well. Real robots are slow and can break, so much of this training happens in simulation. A four-legged robot trained this way in simulation then walked over mud, snow and rubble in the real world.[2] See the sim-to-real transfer explainer.

    RL in randomized simulation has produced well-known legged-locomotion controllers. Examples include the 2020 ANYmal controller[2] and a transformer-based humanoid walking policy that was deployed outdoors without fine-tuning.[8] The leading generalist manipulation models, by contrast, are trained mainly on real robot demonstrations, sometimes mixed with human video and synthetic data.[9][10] RL is now also applied on top of such models in the real world: Physical Intelligence’s Recap method fine-tunes π*0.6 on demonstrations, human corrections and the robot’s own attempts, and the company reports more than doubled throughput on some hard tasks.[11]

    The data bottleneck

    Language models train on text gathered from the internet. Robots have no equivalent source of action data. Roboticist Ken Goldberg calls the difference a “100,000-year data gap”.[4] One response is to pool data. The Open X-Embodiment collaboration gathered more than a million real trajectories from 22 robot types and 21 institutions.[3] Models trained on the mix showed positive transfer across robots.[12] Newer open datasets add variety: DROID collected 76,000 demonstrations in 564 scenes across three continents.[13]

    Companies are also building their own data. NVIDIA trains its GR00T models on real robot data, human videos and synthetic data.[10] Figure reports that pretraining on its human-behaviour dataset raised zero-shot success in new homes from 9% to 56%.[14] In August 2026 Figure turned that into a paid programme, Index, which it says has gathered over 16 million videos from contributors.[15]

    From skills to generalists

    Pooled data and pretrained vision-language models led to generalist policies: one model for many tasks and, ideally, many robots. These are called vision-language-action models (VLAs).[16][12]

    Questions readers ask

    What is imitation learning in robotics?

    The robot learns a policy by copying human demonstrations, often recorded by teleoperating the robot. One 2023 system learned fine two-handed tasks from about ten minutes of demonstrations.[1]

    Why can't robots just learn from the internet like chatbots?

    The internet has huge amounts of text but little data on robot movement. Goldberg calls the difference a "100,000-year data gap".[4]

    Does data from one robot help another robot?

    The Open X-Embodiment project found positive transfer. Models trained on data pooled from 22 robot types improved the capabilities of several robots.[12][3]

    Sources

    Each numbered claim is a statement we checked against the sources listed with it. Status shows how well established it is.

    1. [1]

      A 2023 study using a low-cost bimanual teleoperation system and Action Chunking with Transformers (ACT) learned six fine manipulation tasks at 80-90% success from about ten minutes of human demonstrations. confirmedas of 2023-04-23

    2. [2]

      A 2020 Science Robotics study trained a quadruped (ANYmal) locomotion controller by reinforcement learning in simulation and reported zero-shot transfer to natural terrain such as mud, snow and rubble. confirmedas of 2020-10-21

    3. [3]

      The Open X-Embodiment dataset pooled more than one million real robot trajectories from 22 robot embodiments, collected through a collaboration of 21 institutions and covering 527 skills. confirmedas of 2025-05-14

    4. [4]

      UC Berkeley roboticist Ken Goldberg argued in August 2025 Science Robotics papers that robots face a "100,000-year data gap" compared with the text used to train language models. confirmedas of 2025-08-27

    5. [5]

      The ACT paper identifies compounding errors over time as a key challenge in imitation learning and addresses it by predicting sequences ("chunks") of actions. confirmedas of 2023-04-23

    6. [6]

      Diffusion Policy (2023) represents a robot's visuomotor policy as a conditional denoising diffusion process and reported an average 46.9% improvement over prior methods across 12 tasks from 4 benchmarks. confirmedas of 2023-03-07

    7. [7]

      Physical Intelligence's π0 (October 2024) is a vision-language-action model that adds a flow-matching action generator on top of a pretrained vision-language model. confirmedas of 2024-10-31

    8. [8]

      A 2023 study trained a transformer-based humanoid walking controller with reinforcement learning in randomized simulated environments and deployed it on a real humanoid outdoors without further fine-tuning. confirmedas of 2023-12-14

    9. [9]

      π0 was trained on data from multiple robot types, including single-arm, dual-arm and mobile manipulators, and evaluated on tasks such as laundry folding, table cleaning and box assembly. confirmedas of 2024-10-31

    10. [10]

      GR00T N1 was trained on a mixture of real-robot trajectories, human videos and synthetically generated data, and was deployed on the Fourier GR-1 humanoid. confirmedas of 2025-03-18

    11. [11]

      In November 2025 Physical Intelligence introduced π*0.6, trained with a reinforcement-learning method called Recap that combines demonstrations, corrections and autonomous experience; the company says it more than doubled throughput on some of the hardest tasks. confirmedas of 2025-11-17

    12. [12]

      The RT-X models trained on Open X-Embodiment (paper first posted October 2023) showed positive transfer, improving the capabilities of multiple robots by using experience from other robot platforms. confirmedas of 2023-10-13

    13. [13]

      The DROID dataset (2024) contains 76,000 demonstration trajectories, or 350 hours of interaction, collected across 564 scenes and 84 tasks by 50 collectors in North America, Asia and Europe over 12 months, and is fully open source. confirmedas of 2024-03-19

    14. [14]

      On September 17, 2026 Figure reported that Helix 2.5 performed tidying, towel folding and bed making in 30 Bay Area homes where no data had been collected, with 56% zero-shot success versus 9% without pretraining on its Index human-behavior dataset. confirmedas of 2026-09-17

    15. [15]

      In August 2026 Figure launched Index, an app that pays people to record videos for robot training; Figure said it had paid $15 million to contributors, received over 16 million videos and committed to spend over $1 billion on data and compute in the next 12 months. confirmedas of 2026-08-25

    16. [16]

      Google DeepMind's RT-2 (July 2023) trained a single vision-language model to output robot actions by expressing those actions as text tokens. confirmedas of 2023-07-28

    Revision history (2)
    1. Page created.
    2. Added reinforcement learning on real robots (π*0.6), the DROID dataset and Figure's Index data programme.

    Created Oct 10, 2026. Last reviewed by an editor on Oct 10, 2026. Next scheduled review: Jan 10, 2027.

    Cite this page

    "How robots learn: demonstrations, trial and error, and data." ContentLora, updated Oct 10, 2026. https://contentlora.com/explain/how-robots-learn

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.

    Keep exploring

    ☰ Contents