Source-linked AI summary
A Biophysically Detailed C. elegans Circuit as a Task-Agnostic Dynamical Core for Visually Robust Robot Manipulation
Linrui Qian, Jiajia Zhang, Gan He, Bohan Sun, Zhiwei Lin, Qianhao Wang, Zewu Cai, Nianyu Yi, Mengdi Zhao, Kai Du
TL;DR
Robot policies are typically fitted to a task, body, and visual environment, while whether an evolved nervous-system circuit can supply sensorimotor computation across embodiments remains unresolved. This paper reuses a fixed, biophysically detailed C. elegans circuit as a visuomotor policy’s dynamical core, training only task-specific adapters. The core is competitive across simulated tasks, more robust to visual perturbations than the stated baselines, and its robustness and ablation effects transfer to a real robot.
Problem
Whether a nervous-system circuit detached from its native body can act as a task-agnostic dynamical core for an articulated robot remains an open question.
Source: §1 · p. 2Method
The paper reuses the fixed-weight, biophysically detailed BAAIWorm sensorimotor circuit across tasks, training only task-specific encoder, FiLM, and action-decoder adapters.
Source: §1 · p. 2Results
Across simulated manipulation tasks, the core performs competitively and degrades less under visual perturbation than DP, ACT, and NCP; robustness and ablation effects transfer to a real robot.
Source: §1 · p. 2, §2.4 · p. 5, §3 · p. 6Takeaways & Limitations
The results support a biophysically detailed, biologically constrained sensorimotor circuit as a task-agnostic dynamical core for robot manipulation.
Source: §3 · p. 5Real-robot evaluation covers one manipulation task with a fixed initial pose, and the perturbations do not exhaust distribution shifts; extension to long-horizon planning or dexterous manipulation remains open.
Source: §3 · p. 6
Abstract
from arXiv · showhide
Robot policies are usually trained for one task, one body and one visual environment, and generalize poorly beyond these conditions. Whether a nervous system can instead supply the sensorimotor computation through its evolved wiring and biophysics remains unresolved. Here we embed a biophysically detailed Caenorhabditis elegans sensorimotor circuit - 136 multicompartment neurons with realistic morphologies and electrophysiological characteristics - as the dynamical core of a visuomotor policy. Only thin task-specific adapters are trained; the core's synaptic weights stay fixed while its membrane voltages evolve freely. Across different MetaWorld tasks the core matches or exceeds diffusion-policy, action-chunking-transformer and neural-circuit-policy baselines, and degrades less under visual perturbations. Replacing the core with generic network models such as MLP, LSTM, transformer or reservoir networks removes the advantage. Furthermore, on a real robotic arm the core withstands diverse visual perturbations that collapse the baselines. Our results suggest that visual robustness can be inherited from biophysically detailed circuit dynamics rather than learned by task-specific controllers.
1 Introduction
The paper asks whether a biologically constrained C. elegans circuit can provide a task-agnostic sensorimotor core when detached from its native body. It tests a fixed circuit shared across robot-manipulation tasks, with only task-specific interfaces trained, and reports competitive performance and greater robustness to visual perturbations.
1 Introduction
- The open question is whether a circuit detached from the worm’s body can serve as a task-agnostic dynamical core for an articulated robot.This tests whether evolved wiring encodes a computational prior that survives a change of embodiment.Source: §1 · p. 2
- The approach reuses the complete 136-neuron BAAIWorm sensorimotor circuit with fixed synaptic weights, while training only the encoder, FiLM parameters, and action decoder.RGB images are compressed into sensory drive injected into the circuit as clamp currents; its motor-neuron state is conditioned and decoded into action chunks.Source: §1 · p. 2, Figure 1 · p. 2
- Across simulated manipulation tasks, the fixed C. elegans core performs competitively and degrades less under visual perturbations than DP, ACT, and NCP.Replacing it with MLP, LSTM, transformer, or reservoir cores removes the advantage; robustness and ablation effects also transfer to a real robot.Source: §1 · p. 2
- The paper bounds its claims: it does not claim zero-shot transfer to unseen tasks, generalization across arbitrary bodies or cameras, or exhaustive perturbation coverage.BAAIWorm is reused rather than introduced in this work.Source: §1 · p. 2
2 Results
A fixed-weight, biophysically detailed C. elegans circuit serves as a shared dynamical core, with trained adapters providing task-specific input and output. Across simulated tasks and a real robot, it remains robust under visual perturbations, and matched-interface replacement tests localize the advantage to the circuit.
2.1 A biophysically detailed sensorimotor circuit as a task-agnostic dynamical core
- The 136-neuron circuit has fixed synaptic weights and freely evolving membrane voltages; only the task-specific encoder, FiLM parameters, and decoder are trained.The shared core produces motor-neuron states that are modulated by FiLM and decoded into actions.Source: §2.1 · p. 3
2.2 A single core sustains diverse manipulation tasks under visual perturbation
- Across four MetaWorld tasks, the core matched or exceeded DP, ACT, and NCP clean; under σ = 0.10 noise it retained 0.90–1.00 success versus 0.00–0.26 ACT, 0.08–0.46 NCP, and 0.28–1.00 DP.Under lighting change, core success was 0.82–0.98; under background change, it was 0.80–0.88.Source: Figure 2 · p. 3, §2.2 · pp. 3–4
2.3 Robustness is carried by the circuit, not by the interface
- With the interface held fixed, the circuit’s noisy motor states stayed near their clean manifold, and its speed band was 0.03 versus 0.62 for MLP, 0.28 for LSTM, and 0.35 for transformer replacements.Five trajectory and dynamics metrics reproduced this ordering across the noise sweep.Source: Figure 3 · p. 4, §2.3 · pp. 4–5
2.4 Robustness and ablation effects transfer to a real robot
- On the real robot, the core retained 15–16/20 successes under perturbations, versus 8–14/20 for ACT and 0–12/20 for DP.The largest separations were under illumination change (15/8/0) and object displacement (15/9/6), for core/ACT/DP.Source: Figure 4 · p. 5, §2.4 · p. 5
- On the robot, the full circuit was the only core variant above 50% success under every perturbation; MLP, LSTM, transformer, and reservoir replacements stayed at or below 45%.Figure 4d compares the full core and replacements under nominal conditions and each perturbation.Source: Figure 4 · p. 5, §2.4 · p. 5
3 Discussion
The discussion attributes task-agnostic manipulation performance to the fixed, biologically constrained circuit, with robustness and ablation effects also transferring to a real robot. It connects this result to structured priors and population dynamics while bounding its scope to limited real-robot evidence and a small sensorimotor repertoire.
Main findings
- A biologically constrained circuit supports multiple simulated manipulation tasks, and matched replacements localize its advantage to the circuit rather than adapters, parameter count, or recurrence alone.The reported real-robot results also support transfer of the core's robustness and ablation effects.Source: §3 · pp. 5–6
Connections to prior work
- The results connect low-dimensional population dynamics in neuroscience with structured priors in machine learning, treating the fixed biological circuit as a generator of sensorimotor trajectories.The cited neuroscience literature concerns motor behavior, while the machine-learning connection concerns structured priors and recurrent processing.Source: §3 · p. 6
Limitations
- Real-robot evidence covers one manipulation task from a fixed initial pose, and the tested perturbations do not exhaust distribution shifts.The circuit is reused, and its small sensorimotor repertoire leaves long-horizon planning and dexterous manipulation as open questions.Source: §3 · p. 6
Practical implication
- Using the circuit as the dynamical core confines task-specific training to adapters that map observations into the core and its state into actions.The discussion presents robustness and interpretability as following from the same experimentally constrained object.Source: §3 · p. 6
4 Methods
The methods reuse a fixed-weight, biophysically detailed BAAIWorm circuit as the policy's dynamical core, training only visual and action adapters. They evaluate simulated tasks, perturbations, motor-state dynamics, and a real-robot cup-push protocol.
4.1 Biological core
- The biological core is BAAIWorm's 136-neuron multicompartment model with 1,901 inter-neuron connections; synaptic weights remain fixed while membrane voltages evolve.It includes 1,115 chemical synapses, 786 gap junctions, and 15 sensory-input connections.Source: §4.1 · p. 6
4.2 Interface and training
- Four RGB frames are encoded as a 15-dimensional drive injected into sensory neurons, while FiLM and a decoder use motor-neuron states to predict eight actions, four of which are executed.Behaviour cloning trains the encoder, FiLM parameters, and decoder; the core receives no gradient updates.Source: §4.2 · p. 6
4.3 Baseline policies
- Diffusion policy and ACT use their original recipes, while NCP is a system-level baseline; matched-interface core replacements include MLP, LSTM, transformer, and reservoir models.The matched replacements hold the encoder, FiLM, decoder, optimizer, data, and checkpoint-selection rule fixed.Source: §4.3 · p. 6
4.4 Simulated tasks and perturbations
- MetaWorld evaluation covers coffee-push, hammer, door-open, and assembly, with Gaussian pixel noise (σ = 0.10), illumination changes, and background changes.Success rates use 50 evaluations.Source: §4.4 · p. 6
4.5 Manifold, speed and recovery analyses
- Motor-state analyses measure noisy-trajectory displacement from the clean manifold, speed-profile band width, and five trajectory metrics normalized to clean references.The metrics are temporal jitter, normalized jerk, path curvature, velocity dissimilarity, and tangling.Source: §4.5 · p. 7
4.6 Real-robot protocol and statistics
- A six-degree-of-freedom arm performs cup-push trials from a fixed initial pose under nominal conditions and four perturbations, with 20 trials per condition.An on-site annotator judges success against a fixed criterion; perturbations include image noise, lights off, a green-cloth background, and displaced objects.Source: §4.6 · p. 7