Overview
A robot that folds towels at 97% success can drop to 39% when you add ten steps of Gaussian noise to its actions and flip one gripper command. The same model, with a different action head, barely notices. That contrast is the core of the paper DS-VLA: A Dendritic-inspired Vision-Language-Action Model for Robust Action Control (arXiv 2609.32253, submitted 2026-09-26), which asks not "who scores highest on standard benchmarks" but "who finishes the job after being interrupted."
Key points
An inverted leaderboard
| Model | Nominal success | Perturbed success | Retention | |---|---:|---:|---:| | OpenVLA-OFT | 97.10 | 39.45 | 40.6% | | FAST | 95.20 | 23.90 | 25.1% | | π0 | 94.20 | 28.55 | 30.3% | | DS-VLA | 91.60 | 87.35 | 95.4% | | GR00T | 86.50 | 30.75 | 35.5% |
(Averages over the four LIBERO suites; retention computed as perturbed/nominal.)
- Nominally DS-VLA ranks fourth; under perturbation it ranks first, 47.90 points ahead of the runner-up.
- On the Long suite (long-horizon tasks), DS-VLA drops from 83.4 to 77.8, while the four baselines fall to 1.2, 0.0, 1.0, and 0.0.
- Across difficulty levels, DS-VLA averages 87.35 → 71.65 → 46.00. It degrades; it does not hold indefinitely.
- Three layers, 1536 somatic LIF neurons per layer (learnable thresholds, membrane decay initialized 0.95–0.99).
- Each neuron has two dendritic branches receiving complementary input subsets via fixed binary masks, each with a learnable decay factor—so branches operate at different timescales.
- Readout: running mean of membrane potentials produces the 7-DoF action.
- Training loss: L1 action regression only—no gate loss, firing-rate loss, or dendritic regularization.
- This is a preprint without peer review; author affiliations (Yaxing Lyu, Jingyi Li, Mingkun Xu, Yujie Wu) are not listed on the abstract page.
- Experiments are simulation-only (LIBERO); perturbations are synthetic and don't fully match real mechanical slip, cable tugging, or visual frame drops.
- Baseline numbers are the authors' own reproductions, not original published values.
- DS-VLA's nominal success is still below three baselines—the cost isn't eliminated, just redirected.
- Difficulty 3's 46.00 is still a point on a decay curve; the paper doesn't show where DS-VLA itself goes to zero.
- Yaxing Lyu, Jingyi Li, Mingkun Xu, Yujie Wu, *DS-VLA: A Dendritic-inspired Vision-Language-Action Model for Robust Action Control*, arXiv:2609.32253, 2026-09-26
- LIBERO benchmark (four suites: Goal / Object / Spatial / Long)
- Baselines compared: OpenVLA-OFT, FAST, π0, GR00T
How perturbations are applied
The protocol injects noise into the 7-DoF executed action (translation, rotation, gripper) with no perturbation indicator and no hand-crafted recovery signal:
| Difficulty | Noise std | Steps affected | Extra damage | |---|---|---|---| | 1 | 0.1 / 0.002 | steps 10–19 | gripper command flip | | 2 | 0.010 / 0.004 | 25 steps | 10 translations overwritten with a fixed vector; gripper forced closed 18 steps | | 3 | 0.015 / 0.008 | 35 steps | 12 translations overwritten; gripper forced closed 25 steps |
Difficulty 1 affects only 10 of ~145 steps—under 7% of a trajectory—yet knocks OpenVLA-OFT from 97.1 to 39.45.
Recovery behavior: as difficulty rises from 1 to 3, DS-VLA's average successful trajectory length grows (145.4 → 155.2 → 175.4 steps) while action frequency (~6.5 Hz), inference latency (~120 ms), and closed-loop latency (~150 ms) stay flat. It recovers by taking detours, not by acting harder.
The dendritic spiking action head (DH-LIF)
The intuition: spiking neurons carry state. The membrane potential persists frame to frame, giving the control loop a built-in short memory—"what was I just doing?"—which is the substrate for recovery. Transformer-style action heads re-read everything each frame.
The gate is what matters
Each neuron computes an inhibitory gate g ∈ [0, 1] from the whole layer, shared by its two branches. It scales only incoming current: high g blocks new evidence. Crucially, it does not clear stored dendritic state—the gate decides whether new information is admitted, not whether memory is erased. Noise arrives → gate closes → outliers don't write. Noise passes → gate opens → normal information flows again.
Ablation results (removing the gate):
| Metric | Change | |---|---| | Mean nominal success | −18.0 pts | | Mean perturbed success | −20.9 pts | | Long suite nominal | −48.4 pts | | Long suite perturbed | −42.2 pts | | Object suite perturbed | +0.8 pts (only reversal, likely noise) |
Long-horizon performance depends almost entirely on this gate.
Why it matters
1. The evaluation gap is exposed. Nominal success and closed-loop robustness are nearly two different capabilities in current VLAs. Reporting only nominal scores hides a drop from 97 to 39 that would be fatal on real hardware. 2. A cheap direction. Bigger vision-language backbones and stronger generative action decoders don't fix this—the failure under perturbation is in temporal integration, not perception. A gated, decaying-memory action head costs very little. 3. The gate is portable. It's a layer on the action head, independent of the VLA backbone—applicable to any control setting where observations occasionally go bad.