Accelerated co-design of robots through morphological pretraining¶

Conference: ICLR2026
OpenReview: https://openreview.net/forum?id=WVliGyFwZv
Code: Project website provides videos and code (link on the final page of the paper)
Area: Robotics / Morphological-Control Co-design
Keywords: Co-design, Differentiable Simulation, Morphological Pretraining, Universal Controller, Evolutionary Algorithms

TL;DR¶

This paper introduces "morphological pretraining": a morphology-agnostic universal controller is pretrained once across tens of millions of robot bodies using differentiable simulation. This frozen (or slightly fine-tuned) controller then enables zero-shot evaluation of arbitrary morphological changes, accelerating robot "body+brain" co-design by an order of magnitude and demonstrating, for the first time, that evolutionary "crossover" can produce offspring superior to their parents.

Background & Motivation¶

Background: Robot co-design requires simultaneous optimization of physical morphology and neural control. The prevailing approach uses Reinforcement Learning (RL) to learn a separate control policy for each candidate body, as changing a single morphological part alters the control gradients entirely.

Limitations of Prior Work: RL requires massive interaction data to approximate control policies in non-differentiable simulations. During evolution, morphological changes necessitate relearning controllers repeatedly, leading to a computational explosion. Consequently, the field has struggled for thirty years—most work is limited to exploring thousands of morphologies on "rigid stick-figures with fewer than a dozen parts." While soft robots have many components, they often lack actuators and intelligent sensing, and are frequently restricted to 2D environments.

Key Challenge: The "learning a separate controller for every body" paradigm is the bottleneck: it nests morphology search (outer, discrete, non-differentiable) with control learning (inner, data-intensive), requiring a full training cycle for every step in the outer loop. More subtly, attempting to "simultaneously co-learn a universal controller from scratch" leads to diversity collapse, where the population converges to similar morphologies that are easier for a single shared controller to manage, reducing co-design to "training a controller for a single design."

Goal: (1) Make the controller "immune" to morphology, eliminating the need for retraining on new bodies; (2) Enable instant evaluation of non-differentiable changes like adding, removing, or recombining body parts; (3) Solve/leverage diversity collapse to achieve high performance while maintaining morphological diversity.

Key Insight: Drawing from the success of large-scale pretraining in CV/NLP—if language models can pretrain universal capabilities on massive corpora, why can’t controllers pretrain "universal driving skills" on massive bodies? The key is differentiable simulation: it provides direct gradients for controller parameters, making it feasible to "average gradients across millions of bodies," a scale unattainable for RL due to the lack of gradient information.

Core Idea: Use differentiable simulation to pretrain a morphology-agnostic universal controller on 10M+ bodies to obtain a "prior brain." Subsequently, freeze this controller for zero-shot evolution (instant evaluation of morphological changes) or perform few-shot evolution with minor per-generation fine-tuning, thereby completely decoupling control learning from the inner loop of morphological search.

Method¶

Overall Architecture¶

The method consists of three integrated stages: defining robot encoding and simulation within a unified "body space → physical entity → differentiable environment" pipeline; pretraining a shared universal MLP controller on millions of random bodies, terrains, and light sources to learn how to drive nearly any morphology toward a goal (phototaxis); and finally using this controller as a prior for a standard Genetic Algorithm (GA) to evolve the body population. Freezing the controller enables zero-shot evolution, while resetting and fine-tuning for 60 steps per generation enables few-shot evolution. The pipeline takes "random body genotypes + random environments" as input and outputs "a population of high-performance diverse evolved bodies + a controller that drives them."

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Random Body Genotype<br/>+ Random Terrain + Light"] --> B["Voxel Genotype to<br/>Spring-Mass Phenotype"]
    B --> C["Differentiable Mass-Spring Sim<br/>Phototaxis + Terrain Contact"]
    C --> D["Morphological Pretraining<br/>Universal Controller Shared by 10M Bodies"]
    D -->|Freeze Controller| E["Zero-Shot Evolution<br/>Instant Evaluation of Morphological Changes"]
    D -->|Reset + 60-step Fine-tune| F["Few-Shot Evolution<br/>Resisting Diversity Collapse"]
    E --> G["Diverse High-Performance Population"]
    F --> G

Key Designs¶

1. Voxel Genotype to Spring-Mass Phenotype: Making bodies both evolvable and differentiable

Bodies are encoded as \(6\times6\times4\) binary voxel genotypes \(G\). Each occupied voxel maps to a \(10\,\text{cm}^3\) cubic unit in the physical phenotype \(P\), with masses at the eight corners and springs along the edges and face diagonals. Adjacent voxels share interface masses and springs to ensure structural cohesion. The workspace accommodates up to \(|M|=245\) masses and \(|S|=1648\) springs. This design bridges discrete, bit-flip/XOR-recombining genotypes (suitable for evolution) with continuous mass-spring physical bodies (suitable for differentiable simulation). The authors also handle symmetries (90° rotations, x/y mirroring) using lexicographical normalization to avoid redundant body representations.

2. Morphology-Agnostic Universal Controller: Using input/output masking for any body

The controller is a simple MLP (input: 250 dimensions = 245 mass light sensors + 5 CPG sine waves; three 256-unit hidden layers; output: 1648 dimensions for all possible springs; 620,912 parameters). To handle varying sensor/actuator counts across bodies, the input and output dimensions are fixed to \(|M|\) and \(|S|\) respectively. Sensors or actuators not present in a specific body are masked to zero. This provides "implicit morphological conditioning" through the observation and action space masking; the network learns the body structure by detecting "active" channels. Light readings are zero-centered by subtracting the mean of all active sensors to provide an "embodied irradiance gradient." Springs operate according to Hooke's Law \(F=k(L-L_0)\), where the rest length \(L_0\) is driven within \(\pm20\%\) of its nominal value.

3. Large-Scale Morphological Pretraining: Averaging gradients over 10M+ bodies

The controller is pretrained for 1400 steps across 10 million distinct bodies, where each sample (random body, terrain, and light source) is seen only once. The objective is to minimize the batch mean of \(d_1/d_0\), where \(d_0\) and \(d_1\) are the initial and final distances to the light source. This relative distance formulation incentivizes both far-away and nearby individuals equally. Simulation is implemented in Taichi, performing end-to-end backpropagation over 1000 physical steps (\(dt=0.004\text{s}\)). Gradients are averaged across variations in bodies, worlds, and targets (Batch size 8192 on 8x H100). This scale is unique to differentiable simulation; convergence is reached in 57 minutes, with loss dropping from 1.0 to approximately 0.3 (completing 70% of the initial distance).

4. Zero-Shot / Few-Shot Evolution: Decoupling control from morphological search

With a pretrained prior, evolution focuses purely on morphology. Zero-shot evolution: The controller is frozen while a GA evolves 8192 individuals (25% bit-flip mutation, 75% XOR crossover). Since the controller is static, morphological changes can be evaluated instantly. However, this leads to diversity collapse, producing "clones" optimized for the fixed pretrained model. Few-shot evolution addresses this: at the start of each generation, the controller weights are reset to the pretrained values, and the optimizer is reset before fine-tuning for 60 steps on the current population (30 for parents, 30 for offspring). This "reset-and-slight-finetune" strategy surprisingly maintains and significantly increases diversity while achieving superior performance, as the controller adapts to the "current population" rather than forcing the population to accommodate a fixed controller.

Loss & Training¶

The pretraining loss is the batch mean of the relative distance ratio \(d_1/d_0\). Optimization uses Adam (\(\beta_1=0.9, \beta_2=0.999\), gradient clipping 1.0) with cosine annealing learning rate restarts (initial \(1\mathrm{e}{-3}\), minimum \(1\mathrm{e}{-5}\), period doubling after each restart). Few-shot fine-tuning uses \(3.5\mathrm{e}{-4}/3.5\mathrm{e}{-5}\) with a period of 100 truncated at 60 steps.

Key Experimental Results¶

Main Results¶

Performance and diversity comparison across three co-design paradigms (Diversity = normalized mean pairwise Hamming distance in genotype space):

Paradigm	Pretrained	Convergence (Steps/Time)	Performance	Diversity
Morphological Pretraining	—	1400 steps / 57 min	Loss 1.0→0.3 (70% Gain)	Covers heterogeneous bodies
Zero-Shot Evolution	Yes (Frozen)	100 gen / 17 min	Rapidly nears optimum	Collapse (Clones)
Few-Shot Evolution	Yes (Reset+60 steps)	~18 gen / 53 min	Best and continuously improving	Significant increase & sustained
Simultaneous Co-design (Li et al. 2025)	No	<180 gen / 109 min	Similar training loss	Rapid collapse

Key takeaway: Few-shot evolution at Gen 6 (G6) and zero-shot at Gen 31 (G31) already match or exceed the performance of Simultaneous Co-design at Gen 180 (G180).

Ablation Study¶

Simultaneous co-design serves as an ablation by removing pretraining and resetting:

Configuration	Performance	Diversity	Note
Few-Shot (Full)	Best	Significantly increases	Pretraining + Reset-finetune
Zero-Shot (No Finetune)	Rapid near-optimum	Collapse	Pretrained only
Simultaneous (No Pretrain/Finetune)	Slower, needs G180	Rapid collapse	Prev. SOTA (Li et al. 2025)

Key Findings¶

Diversity collapse is an inherent pathology of co-design: When a population adapts to a shared controller (frozen or learned from scratch), evolution collapses to a single "species."
"Reset-and-finetune" is the cure: By forcing the controller to adapt to the current population, diversity increases spontaneously without explicit selection pressure.
Pretrained controllers unlock effective crossover: In zero-shot evolution, improvements in offspring can be attributed solely to morphological recombination, providing rare evidence that "recombination produces superior offspring" in robot evolution (Fig 5: Parent loss 0.257/0.593 → Offspring 0.073).
Robustness & Generalization: The controller remains functional when failing 1/4 of motors or over half of sensors. Zero-shot evolution can adapt bodies to discrete platforms or magnetic-field targets without fine-tuning the controller.
Emergence of Saltation: The universal controller discovered a kangaroo-like saltation (jumping) gait with distinct flight phases, which differs from gaits typically designed for single bodies.

Highlights & Insights¶

Applying the "Pretraining Paradigm" to Co-design: The core insight is that relearning controllers is the bottleneck; differentiable simulation enables "universal pretraining" that is unattainable with RL.
Masking as Morphological Conditioning: Masking inputs/outputs allows a fixed-dimension MLP to control thousands of bodies, suggesting that many "custom-net" scenarios can be replaced by shared backbones + masking.
Identifying a New Pathology: Characterizing "diversity collapse" and providing a simple "reset-to-pretrained" cure is a major contribution.
Directionality of Adaptation: Whether the population adapts to the controller or vice versa determines the survival of diversity.

Limitations & Future Work¶

Ours considers single material (soft), single sensing (light), single actuator (linear springs), and single task (phototaxis). Future work should address multi-task/multi-material scenarios.
Sim-to-real: Experiments were conducted only in simulation. Real-world transfer may require higher resolution or noise modeling.
Dependency on Differentiable Simulation: The method requires an analytical gradient, which may not be available for all contact-heavy or fluid environments.

Vs. Simultaneous Co-design (Li et al. 2025): They inherit controllers across generations; Ours uses large-scale pretraining. The key difference is adaptation direction—theirs forces the population to adapt to the controller (causing collapse), while Ours forces the controller to adapt to the population (preserving diversity). Ours achieves G180-level performance by G6 or G31.
Vs. RL General Controllers (MetaMorph, etc.): RL methods approximate universal policies on small, pre-designed sets of morphologies. Ours leverages differentiable simulation to pretrain on a scale (10M+) impossible for RL.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ Introduces pretraining to co-design and solves diversity collapse with a novel paradigm.
Experimental Thoroughness: ⭐⭐⭐⭐ Extensive comparisons and robustness tests, though limited to simulation.
Writing Quality: ⭐⭐⭐⭐⭐ Very clear logic and effective visualization of mechanisms.
Value: ⭐⭐⭐⭐⭐ Significantly improves efficiency and scale; the methodology is highly transferable.