Vision-language-action models decide at the speed of a conscious thought. Bodies get hit at the speed of a spinal reflex. SPINE (codename Sandevistan) is a small, embodiment-agnostic reflex layer that lives between the VLA's plan and the actuators: it reacts in milliseconds, slows the plan's clock when it must, and overrides it when it has to.
Named after the Sandevistan of Cyberpunk 2077, a spinal implant that lets its owner move while the world stands still, and after the bullet-time shots of The Matrix. In the body the trick is older than either: it is the spinal cord.
A conscious reaction to something you see takes about 200 ms. A spinal reflex has acted in 20–40 ms, the eye-stabilising reflex in about 10 ms, and a fish has already turned away from a predator in the same time. These are not the brain being fast; they are loops that do not go through the brain at all. The cortex sets goals and tunes the reflexes, the spinal cord and brainstem execute, and when the two disagree in an emergency the reflex wins first and the cortex finds out afterwards.
A tap stretches the quadriceps; muscle spindles fire, one synapse in the spinal cord, and the muscle contracts about 30 ms later. The same loop keeps posture up against every small push, all day, without a thought. Lesson: a one-hop loop with the right sensor beats a smart loop that is late.
Image: Amiya Sarkar, Wikimedia Commons, CC BY-SA 4.0.
Step on a nail: the leg lifts before the pain is felt, and the other leg stiffens so you do not fall. One stimulus, a coordinated whole-body response, all wired below the brain. Lesson: reflexes are not single-joint; they know the body.
Image: Cenveo / OpenStax Anatomy & Physiology, Wikimedia Commons, CC BY 4.0.
Turn your head and the eyes counter-rotate within about 10 ms, driven by the inner ear's accelerometers and gyroscopes. The image stays still while you move. Lesson: the IMU is the fastest sense there is, and it belongs in the reflex, not in the planner.
Image: .Koen, Wikimedia Commons, CC BY-SA 4.0.
A puff of wind on a cockroach's cerci reaches giant interneurons and it is turning away in tens of milliseconds. A fish's Mauthner cell triggers a C-start in about 10 ms. These circuits take over whatever the animal was doing. Lesson: an override path with the authority to interrupt the plan.
Image: Nickfox14, Wikimedia Commons, CC BY-SA 4.0.
Brush the top of a walking cat's paw while the leg swings forward and the leg lifts higher to clear the obstacle. Brush it while the paw is planted and the leg pushes down harder. Same stimulus, opposite response, chosen by the phase of the ongoing movement (Forssberg, 1979). Lesson: a good reflex reads the plan's phase before it acts. This is the biological version of the "bullet time" rule below.
Press Start, wait for the pad to light up, then click (or tap) it as fast as you can. Your time is compared with a spinal reflex and with a VLA that has to finish one inference and start a new chunk.
A VLA such as π0.5 or GR00T N1 reads language and images, thinks for 100–300 ms, and emits a chunk of actions that the robot then plays back for a second or more. Everything that happens to the body inside that window, a gust, a slip, a bump, a payload shift, is handled by whatever low-level controller sits underneath, which knows nothing about the task. SPINE is the missing middle: a small network that runs at hundreds of hertz on proprioception and fast visual cues, sees the VLA's chunk and its intent, and decides in each millisecond whether to let the plan run, correct it, slow it down, or take over.
The obvious way to override a plan is to replace its actions. That throws the plan away and leaves the VLA confused about where its chunk went. The cat's stumbling reaction suggests a better rule: act relative to the phase of the plan. SPINE owns the phase variable that plays the VLA's chunk. Under a small disturbance it adds a residual and keeps the clock running. Under a large one it slows the clock to a crawl, corrects the body while the plan "waits", then lets the clock catch up. The chunk is never discarded, so when the disturbance is over, the robot is exactly where the plan expects it to be, just a little later. Dynamic movement primitives have used phase stopping for perturbations for two decades; here the phase belongs to a learned reflex and the plan comes from a foundation model.
The magenta dot is where the plan says the robot should be right now; the cyan one is the robot. Hit it with a gust and watch what each mode does with the plan's clock. Nothing here is learned, it is the control rule, drawn.
Reflex circuits are the most conserved part of the vertebrate nervous system; what differs between a fin and a leg is the wiring to the muscles. SPINE copies the split. Its inputs and outputs are body-level quantities every robot has, body twist, acceleration, contact, the plan's phase, and a thin adapter per robot maps its output onto that robot's controller: wrench targets for a multirotor's allocator, foot forces for a quadruped's MPC, joint velocities for an arm. The reflex is pre-trained across several bodies and a new one gets an adapter and a short fine-tune, the way a VLA gets a new action head.
Within 2025 every major robot-learning group moved to a two-system design: a slow vision-language planner and a fast action generator. That is the cortex and the brainstem. None of them has a spinal cord yet: a layer that is independent of the body, runs at reflex speed on reflex senses, and has the authority to override.
"System 2", a 7B-parameter VLM at 7–9 Hz, produces a latent that "System 1", an 80M-parameter visuomotor transformer, turns into whole-upper-body commands at 200 Hz. The two run as separate processes; S1 observes the robot at a higher rate than S2. figure.ai/news/helix
Closest in spirit. S1 is trained for one humanoid and does not override S2; it executes it.
System 2 is an Eagle-2 VLM at 10 Hz; System 1 is a diffusion transformer that "generates closed-loop motor actions at a higher frequency (120 Hz)" with 16-step chunks, trained end to end across human video, synthetic and real robot data. arXiv:2503.14734
Fast system, but co-trained with the slow one and tied to the embodiment heads; no separate reflex authority.
Treats asynchronous chunk generation as inpainting: the next chunk is produced while the current one plays, frozen where the robot is already committed, so π0.5 keeps moving through 300 ms of inference latency. arXiv:2506.07339 · pi.website
Solves the pause between chunks, not the chunk being wrong after a push. Complementary: SPINE would sit under it.
Gemini Robotics-ER 1.5 plans and reasons, Gemini Robotics 1.5 acts and can "think before acting"; motions learned on ALOHA 2 transfer to Apptronik's Apollo and a bi-arm Franka. The On-Device model runs locally for latency-sensitive use. DeepMind blog · On-Device
Cross-embodiment transfer at the planning level; still a VLA-speed loop.
One model directly controls the whole humanoid, hands and feet alike, through a long sequence of handling and walking; researchers push boxes around mid-task and Atlas adjusts. TRI press release
Whole-body learning at policy rate; the reflex-speed recovery still comes from Boston Dynamics' underlying controllers.
This is a research vision, not a result. Nothing described in Parts 2–4 has been built; the data pipeline, the disturbance models, the specialists and the fully verified dataset that Step 2 relies on exist in the aerial manipulation project. I am writing it down because I think it is the right next problem, and because I would like to spend a PhD on it. If you run a lab where this belongs, I would be glad to talk.
Mingtian Tan · mingtian.tan20@gmail.com. All illustrations in Part 1 are from Wikimedia Commons under their own licences (credits under each image; full list in assets/media/CREDITS.md); the figures in Parts 1–2 are drawn for this page. The two clips under the title are streamed from GIPHY and belong to Warner Bros. and to Studio Trigger, CD PROJEKT RED and Netflix. Sandevistan is a name from Cyberpunk 2077 by CD Projekt Red, used here as a nickname.