A legged robot is an awkward subject for trial-and-error learning. It falls, and falls damage it. Its dynamics are hybrid, switching at every footfall, and the forces that matter most pass through contacts that are difficult to measure. Reinforcement learning needs a volume of experience that no physical machine can supply in reasonable time. Simulation supplies it cheaply, and the price is the mismatch between the simulated and the physical robot. The term for that mismatch, the reality gap, comes from evolutionary robotics, where Jakobi et al. (1995) observed that controllers evolved in simulation transferred only when the simulation contained appropriate noise.

For many years the strongest legged controllers were model based. Whole-body and model predictive control gave quadrupeds such as the MIT Mini Cheetah fast and dynamic gaits (Kim et al., 2019), and learned policies were a laboratory curiosity by comparison. That changed quickly. Within a few years policies trained entirely in simulation were shown trotting and galloping on a small quadruped (Tan et al., 2018), recovering a larger one from falls (Hwangbo et al., 2019) and carrying it over rough natural ground (Lee et al., 2020).

This review asks two questions of that literature. Which ideas account for the progress, and which parts of the gap remain open once the demonstrations are set aside? The second question receives the more attention, because the successes are well publicised and the failures are seldom written up.

The article is a narrative review of the published literature on transferring learned locomotion controllers from simulation to legged hardware. It concentrates on reinforcement learning for quadrupeds and bipeds, and it draws on peer-reviewed articles, conference papers and widely cited preprints known to the field. It does not attempt an exhaustive or systematic search, and it does not rank methods numerically, since the reported experiments differ in robot, terrain and measure. General surveys of sim-to-real methods in robotics are available elsewhere (Zhao et al., 2020; Muratore et al., 2022); the present review is narrower and limited to legs.

Figure 1 groups the sources of mismatch, the remedies and the open problems that organise the rest of this review. Four sources recur in the literature.

The remedies fall into five families, which are usually combined. Table 1 summarises them with representative sources.

The evidence for these methods consists of hardware demonstrations. On quadrupeds the sequence runs from flat-ground gaits (Tan et al., 2018) through fall recovery and fast running (Hwangbo et al., 2019) to blind traversal of mud, snow and rubble (Lee et al., 2020), perceptive hiking on an alpine trail (Miki et al., 2022) and parkour-style climbing and jumping (Hoeller et al., 2024). Bipeds, with a smaller support polygon and less margin, followed. Policies for the Cassie biped were transferred after iterative refinement of the training design (Xie et al., 2020) and later climbed stairs without vision (Siekmann et al., 2021). Radosavovic et al. (2024) reported a full-size humanoid walking outdoors under a policy trained only in simulation.

These results are persuasive, and they should be read for what they are. Each is a demonstration on one robot, selected by its authors, in conditions they chose. Success rates are reported by some groups and not by others, and negative results are rare. The literature shows that transfer can work. It says much less about how often it does, or by what margin.

Terrain that deforms or moves. Training terrains are rigid height fields. Sand, deep snow, vegetation, loose scree and floating or swaying platforms violate that assumption. Blind policies cope with some of these through adaptation, but they do so by extrapolation, and no widely used simulator models granular or yielding ground at the speed that learning requires.

Perception. Policies that use cameras or lidar are trained on rendered depth or sampled height maps with synthetic noise. The failure modes of real sensors are structured, not random, and correlated with the very situations in which the robot most needs to see. The belief-state design of Miki et al. (2022) mitigates the problem by falling back on proprioception; it does not make the simulated sensor realistic.

Evaluation. There is no accepted test course, no standard perturbation set and no common reporting format for sim-to-real locomotion. Papers report different robots on different ground with different measures. Whether one method transfers better than another is in most cases impossible to establish from the published record, a weakness that surveys of the wider field also note (Muratore et al., 2022).

Reward and curriculum engineering. Policies that transfer well are trained with long lists of shaped reward terms that penalise foot slip, joint acceleration, torque and action rate. These terms are partly a disguised remedy for the gap, since they discourage behaviour the hardware cannot reproduce, and they are tuned by hand against the real robot. The knowledge involved is rarely documented in full.

Hardware condition and safe failure. A learned actuator model describes a robot on the day the data were logged. Gearboxes wear, batteries sag and feet are replaced. Little published work examines how transferred policies degrade as hardware ages, or how a policy should behave when the situation lies outside anything it met in training.

Learning on the robot itself is one response. Haarnoja et al. (2019) trained a small quadruped to walk in the real world in about two hours, Smith et al. (2022) reduced this to around twenty minutes with a carefully tuned off-policy method, and Wu et al. (2023) used a learned world model for the same purpose. Such methods avoid the gap by never entering simulation, but so far for simple gaits on benign ground. The likelier role for real-world learning is to refine a policy pre-trained in simulation.

Three lines of work would address the gaps identified above. The first is measurement. A shared set of physical test conditions, with agreed perturbations and reporting of failures as well as successes, would let the field compare methods and would expose which sources of mismatch still matter. The second is simulation of what is currently omitted: compliant and granular ground, sensor models built from recorded failure cases, and actuator models that include temperature and wear. The third is a principled account of the residual gap at run time, so that a policy can recognise when its training no longer applies and hand over to a conservative fallback.

A further question concerns the division of labour between learning and models. Model-based controllers come with constraints that can be checked, and learned policies with robustness that cannot yet be certified. Hybrid schemes, in which a learned policy proposes and a model-based layer verifies, are an obvious meeting point and are likely to matter most where legged robots work near people.

Sim-to-real transfer for legged locomotion has moved from doubtful to routine through a small number of ideas applied together: honest actuator models, randomised dynamics, adaptation from proprioceptive history, privileged training and very fast simulation. The demonstrations that followed are real achievements. They are also selected ones, and the published record does not yet support quantitative claims about reliability. Deformable terrain, realistic perception, comparable evaluation, undocumented reward engineering and ageing hardware are where the gap persists. Progress on those fronts depends less on new learning algorithms than on better measurement and on simulators that include what has so far been left out.

To be declared by the author(s) before publication.

To be declared by the author(s) before publication.