SMOL-Human Schedule for Robot Locomotion

• SMOL-Human Schedule is a biologically inspired curriculum that modulates robot actuator strength in phases mirroring human developmental power dynamics.
• It integrates with MAP-Elites by re-evaluating solution archives to filter behaviors and ensure transferability from exploratory high-torque phases to stable adult conditions.
• Experimental results on benchmarks like HalfCheetah and Walker2D demonstrate notable improvements in speed and coverage compared to static baseline methods.

Searching arXiv for the cited papers to ground the article in the relevant sources. SMOL-Human Schedule is a biologically inspired actuator-scaling curriculum introduced in “Time to Play: Simulating Early-Life Animal Dynamics Enhances Robotics Locomotion Discovery” (Templier et al., 15 Sep 2025). Within that work, it denotes a non-linear torque schedule for simulated robots that models isometric body variations due to non-linear changes like puberty, with the explicit aim of improving the discovery of locomotion behaviours in quality-diversity search. The schedule is embedded into MAP-Elites by varying actuator torque across training, so that agents experience a low-power childhood, a higher-power adolescence and early adulthood, and then a return to standard adult torque for final evaluation under identical adult conditions (Templier et al., 15 Sep 2025).

1. Definition and conceptual role

In (Templier et al., 15 Sep 2025), SMOL-Human is a schedule over actuator strength rather than a change to morphology, controller architecture, or archive structure. It belongs to the broader SMOL family, short for Scaling Mechanical Output over Lifetime, which dynamically modulates robot actuator strength to mimic natural variations in power-to-weight ratio during growth and ageing. The specific SMOL-Human variant is based on studies of the total power output in humans and is intended to capture the fact that humans’ muscle power-to-weight ratio does not rise or fall monotonically.

The schedule is motivated by a developmental analogy. After a low-power childhood, total power per kilogram increases sharply through puberty, peaks in early adulthood, and then gradually declines with age. The paper links this developmental trajectory to play and exploratory locomotion in juvenile animals, arguing that periods of relatively high muscular strength can help discover diverse gaits that later serve as stepping stones when the adult body emerges (Templier et al., 15 Sep 2025). In the robotics formulation, this becomes a curriculum for exploration and subsequent adaptation rather than a literal model of human ontogeny.

A potential source of confusion is terminological. The label “SMOL-Human” is also used in “Human-Robot Team Coordination with Dynamic and Latent Human Task Proficiencies: Scheduling with Learning Curves” (Liu et al., 2020), where it expands to “Stochastic, Multi-objective Optimization with Learning curves for Humans” and refers to a human-robot task-scheduling framework. In current arXiv usage, the same string therefore names two distinct constructs: a locomotion curriculum in (Templier et al., 15 Sep 2025) and a resource-coordination framework in (Liu et al., 2020).

2. Biological inspiration and curriculum hypothesis

The biological premise in (Templier et al., 15 Sep 2025) is explicitly non-monotonic. Rather than assuming a static actuator model or a simple linear change in strength, SMOL-Human follows a “bell-shaped” trajectory: weak in childhood, stronger through puberty, then lower again under adult conditions. This contrasts both with fixed-$\alpha$ training and with a simple linear decay schedule.

The central hypothesis is stated in developmental and search-theoretic terms. By mimicking this non-linear trajectory in simulated robots, controllers may leverage an early-life high-power phase to explore unusual behaviours, then adapt those behaviours to the mature body under more constrained torque, yielding greater final performance and diversity than training with static actuators or a simple linear decay (Templier et al., 15 Sep 2025). The schedule is therefore not merely a domain randomization scheme. Its phases are ordered and interpretable: the rise phase expands accessible behaviours, the decline phase filters behaviours that depend on elevated torque, and the final adult segment standardizes comparison.

This suggests that SMOL-Human functions as a structured developmental curriculum for QD search. The data in (Templier et al., 15 Sep 2025) make this interpretation explicit by connecting early high-torque exploration to “stepping stones” and the later decline to adaptation under the target adult physics.

3. Mathematical specification

Training is divided into $P=100$ equal “phases” (batches of generations) (Templier et al., 15 Sep 2025). Let $\phi \in [0,1]$ denote the normalized phase index: $\phi = \frac{\text{current phase number}}{P}.$

The torque scaling factor $\alpha(\phi)$ is defined piecewise as

$\alpha(\phi) = \begin{cases} \alpha_{\rm child} + (\alpha_{\rm peak}-\alpha_{\rm child})\dfrac{\phi}{\phi_{\rm peak}} & 0 \le \phi \le \phi_{\rm peak}, \[6pt] \alpha_{\rm peak} - (\alpha_{\rm peak}-\alpha_{\rm adult})\dfrac{\phi-\phi_{\rm peak}}{\phi_{\rm adult}-\phi_{\rm peak}} & \phi_{\rm peak}<\phi\le \phi_{\rm adult}, \[6pt] \alpha_{\rm adult} & \phi_{\rm adult}<\phi\le1, \end{cases}$

with the biologically inspired parameters

$$\alpha_{\rm child}=0.7,\quad \alpha_{\rm peak}=1.4,\quad \alpha_{\rm adult}=1.0,\quad \phi_{\rm adult}=0.90.$$

The schedule is described in words in three segments (Templier et al., 15 Sep 2025):

• From phase $0$ to $\phi_{\rm peak}$ (end of adolescence), torque is linearly ramped from $0.7$ to $P=100$0.
• From $P=100$1 to $P=100$2, torque is linearly declined from $P=100$3 back to $P=100$4.
• For the last 10 phases, that is $P=100$5, $P=100$6 is fixed so that all methods are compared under identical adult conditions.

Two features are central here. First, the peak torque exceeds the adult torque, so exploration is temporarily facilitated under conditions not retained at the end of training. Second, the final segment enforces direct comparability, because all methods terminate under the same adult actuator setting (Templier et al., 15 Sep 2025).

4. Integration into MAP-Elites

SMOL-Human is integrated into CVT-MAP-Elites rather than treated as a preprocessing stage. The implementation in (Templier et al., 15 Sep 2025) uses an archive with 1 024 cells, and at each phase the simulator’s actuator gear ratio is updated according to the current $P=100$7. The entire archive is then re-evaluated under the new torque, after which evolution proceeds for $P=100$8 generations under that same phase-specific physics.

The pseudocode in (Templier et al., 15 Sep 2025) specifies the loop as follows in operational terms. The archive $P=100$9 is initialized with random solutions and $\phi \in [0,1]$0. For phase $\phi \in [0,1]$1, one computes $\phi \in [0,1]$2, applies the piecewise definition of $\phi \in [0,1]$3, updates actuator torque, re-evaluates every individual in the archive under the new physics, and then performs mutation and crossover for the phase’s allotted generations before moving to the next phase. The returned object is the final archive $\phi \in [0,1]$4, now adapted to adult torque.

This design differs from a static-environment QD run in an important way. Re-evaluation of the entire archive after each torque change means that the archive contents are continuously filtered by the current physics. The paper’s own interpretation is that the decline phase “forces the archive to adapt and prune behaviours that rely on superhuman strength, focusing search on those transferable to adult conditions” (Templier et al., 15 Sep 2025). A plausible implication is that SMOL-Human couples exploration and transfer within the archive itself, rather than relying solely on mutation to rediscover adult-feasible behaviours.

5. Experimental setting and quantitative findings

The experimental setup in (Templier et al., 15 Sep 2025) is specific. Simulations use Brax v2 with a JAX-based spring physics backend. The benchmark tasks are HalfCheetah, Walker2D, and Ant Omni. HalfCheetah and Walker2D maximize forward velocity over 1 000 steps, while Ant Omni minimizes energy while moving anywhere in the $\phi \in [0,1]$5-plane in 100 steps, with behaviour descriptor equal to final $\phi \in [0,1]$6. Controllers are feed-forward neural networks with architectures $\phi \in [0,1]$7 for HalfCheetah and Walker2D and $\phi \in [0,1]$8 for Ant. The QD algorithm is CVT-MAP-Elites with 1 024 cells and iso+line mutation using $\phi \in [0,1]$9 and $\phi = \frac{\text{current phase number}}{P}.$0. The total budget is $\phi = \frac{\text{current phase number}}{P}.$1 evaluations with 7–10 random seeds (Templier et al., 15 Sep 2025).

The baselines are also explicitly enumerated in (Templier et al., 15 Sep 2025): Vanilla MAP-Elites with constant $\phi = \frac{\text{current phase number}}{P}.$2, SMOL with simple biomechanical linear decay $\phi = \frac{\text{current phase number}}{P}.$3, SMOL-Reverse $\phi = \frac{\text{current phase number}}{P}.$4, Random-$\phi = \frac{\text{current phase number}}{P}.$5 with phase-wise uniform sampling in $\phi = \frac{\text{current phase number}}{P}.$6, and Extinction ablations removing 5%, 10%, or 50% of the archive before reevaluation under static $\phi = \frac{\text{current phase number}}{P}.$7.

The principal quantitative outcomes reported for SMOL-Human are summarized below.

Task / metric	Vanilla	SMOL-Human
Ant Omni final median coverage	39 cells / 1 024	63 cells
HalfCheetah final median max fitness	7 857	9 123
Walker2D final median max fitness	3 103	3 610

For Ant Omni, the final median coverage rises from 39 cells to 63 cells, corresponding to $\phi = \frac{\text{current phase number}}{P}.$8 coverage with Mann–Whitney $\phi = \frac{\text{current phase number}}{P}.$9 (Templier et al., 15 Sep 2025). For HalfCheetah, final median max fitness rises from 7 857 to 9 123, a $\alpha(\phi)$0 speed improvement with $\alpha(\phi)$1. For Walker2D, final median max fitness rises from 3 103 to 3 610, a $\alpha(\phi)$2 velocity improvement with $\alpha(\phi)$3 (Templier et al., 15 Sep 2025).

The paper further states that, in all three tasks, SMOL-Human significantly outperforms the static baseline and the unstructured Random-$\alpha(\phi)$4 and Extinction ablations. Compared to the simpler SMOL schedule, SMOL-Human is reported as slightly less performant, but it still improves over the static baseline in every metric. The plots reproduced in Fig. 4 show that coverage and max fitness climb above vanilla only after the rise phase and then adapt through the decline phase into the adult constant-torque segment, which the paper interprets as evidence for effective reuse of early-life behaviours as stepping stones (Templier et al., 15 Sep 2025).

6. Interpretation, limitations, and relation to earlier “SMOL-Human”

The interpretation advanced in (Templier et al., 15 Sep 2025) is organized around two insights. First, the early-life torque boost from $\alpha(\phi)$5 to $\alpha(\phi)$6 drives exploration of unusual gaits, and puberty-peak behaviours become “islands of competence” around adult torque. Second, the subsequent torque decline prunes behaviours that depend on elevated strength and retains those transferable to adult conditions. The broader implication proposed in the paper is that play and developmental power-to-weight dynamics in animals may have evolved not only for immediate survival but also as natural curricula that scaffold locomotor learning.

The principal limitation identified in the same source is that only actuator strength is varied. No morphological or limb-length changes accompany the schedule. The paper therefore notes that more realistic ontogenetic models, such as limb growth, may amplify benefits, and it points to more complex curricula, including multiple growth spurts or injury recovery, as future directions (Templier et al., 15 Sep 2025). These are presented as prospective extensions rather than established results.

Because the term “SMOL-Human” also appears in (Liu et al., 2020), it is important to distinguish the locomotion schedule from the earlier scheduling framework. In (Liu et al., 2020), SMOL-Human “weaves together” an exponential-curve-plus-noise model of human learning updated via an adaptive Kalman filter, a joint-chance-constrained makespan-plus-entropy objective over task schedules, a Gaussian upper-bound for probabilistic robustness, and a population-based search policy that balances exploration of agent-task pairings with exploitation of current skill estimates. Its empirical domain is human-robot collaboration: a pilot for learning-curve priors with $\alpha(\phi)$7, a between-subjects human-subject study with $\alpha(\phi)$8, and outcomes including improved team fluency and team efficiency under exploration-favoring scheduling strategies (Liu et al., 2020). Despite the lexical overlap, that framework concerns resource coordination for human teammates, whereas SMOL-Human Schedule in (Templier et al., 15 Sep 2025) concerns actuator-torque curricula for simulated robot locomotion.

Taken together, these two uses of the same acronym reflect distinct research programs. In (Templier et al., 15 Sep 2025), SMOL-Human is a developmentally structured torque schedule within quality-diversity optimization. In (Liu et al., 2020), SMOL-Human is a stochastic multi-objective scheduler for human-robot teams. The former is centered on locomotion discovery through staged mechanical output; the latter on task allocation under dynamic and latent human task proficiencies.