ReflexArc: Bottom-Up Integration of Instantaneous Tactile Reflexes into Whole-Body Humanoid Control

Mingyu Wang, Si Li, Ziqing Chen, He Wang, Li Yi

November 2026

The G1 withdrawing from contacts made by a person.

TL;DR A full-body tactile suit on a Unitree G1 triggers whole-body withdrawal reflexes that reach the motors within 10 ms, and the learned whole-body policy is trained to keep balance without cancelling them. Compared with the same reflex delayed by 20 ms, the first force peak drops by 34.1% and the contact impulse by 2.15x.

What they do

Learned humanoid controllers run at 50 Hz, so a contact can wait up to 20 ms before the policy responds, and a tracking policy treats a protective withdrawal as an error to correct. ARC (Ascending Reflex Control) adds a tactile reflex that bypasses the policy and commands the motors directly, then trains the whole-body policy to keep the robot balanced and on task while the reflex acts on the same joints.

How it works

Tactile startle detection. Piezoresistive fabric patches (128 taxels each, 300 Hz) cover the wrists, elbows, knees, chest, back and feet. Per-taxel adaptive thresholds and connected-component grouping flag a hazardous contact on its first frame, within 5 ms of onset.
Instant reflex solver. A constrained whole-body optimization picks a withdrawal pose that moves the struck point back along the surface normal, charges motion by the whole-body inertia, keeps the support frames fixed and avoids self-collision. A short sequence of local QPs solves it onboard in under 4 ms.
Reflex-aware policy. A PPO motion-tracking policy is trained in Isaac Lab with reflexes injected. The reflex offset is subtracted from its observed joint positions and added back to its outputs, with DAgger-style supervision toward the reflex-free action, so it does not undo the reflex.
Recovery. During a reflex, an AMP reward from human repositioning motions makes the policy walk away from the contact.

Results

Box strike on hardware: against the same reflex delayed by 20 ms, the first force peak drops from 21.64 N to 14.25 N (34.1%) and the impulse from 0.2632 to 0.1225 N s.
Ablations: without reflex-in-the-loop training, survival falls to 82.0% and the robot re-hits the obstacle in 90.9% of trials. Without transparency the policy opposes the reflex and sustained withdrawal drops from 11.06 to 6.74 cm. Tracking an explicit retreat motion instead of AMP reaches 1.5 m in 14.1% of trials versus 63.0%.
Demonstrated on more than ten tracking tasks in clutter, on teleoperation through the ScaleBFM general tracker without retraining, and on stepping over a threshold and climbing a step without vision, using the foot patches.

Limitations

Patches cover only the convex regions that collide most often, so contacts elsewhere go unprotected. The hardware impact numbers come from one controlled box-strike setup. The tracking, teleoperation and traversal cases are reported as demonstrations, without success rates.

Videos

Whole-body tactile reflexes on the arm, hand and knee, with the patch readings shown.

ARC versus a top-down reflex striking a box, slowed 50x, with the contact force traces.

Figures

From Figure 2: Overview of ARC. A hazardous contact is localized within milliseconds, a whole-body reflex pose is sol... — From Figure 2: Overview of ARC. A hazardous contact is localized within milliseconds, a whole-body reflex pose is solved, and the motors withdraw within 10 ms of the collision. The reflex then propagates bottom-up to the policy, which sustains it while continuing its task and draws on a human prior to retreat from the contact region.

From Figure 3: Reflex poses solved for different contacts. From the same initial configuration, the solver produces a... — From Figure 3: Reflex poses solved for different contacts. From the same initial configuration, the solver produces a different whole-body reflex depending on which body region is struck and from which direction (spheres).

From Figure 6: Effect of latency on contact impact. Blue/red: 0/20 ms added delay. Solid: contact force; dashed: hand... — From Figure 6: Effect of latency on contact impact. Blue/red: 0/20 ms added delay. Solid: contact force; dashed: hand displacement. Removing the 20 ms delay reduces the first force peak from 21.64 N to 14.25 N and the contact impulse from 0.2632 N s to 0.1225 N s (2.15x).

Abstract

Humanoid robots operating in cluttered environments must react to unexpected contacts before the first impact peak, which arrives within tens of milliseconds. Learned whole-body policies running at 50 Hz introduce a policy-cycle delay and may suppress a protective withdrawal as a tracking error. To address these challenges, we propose ARC (Ascending Reflex Control), a bottom-up framework that integrates instantaneous tactile reflexes with learned whole-body control. ARC consists of two components: an instantaneous reflex pathway and bottom-up reflex integration. The reflex pathway detects hazardous contacts from full-body tactile sensing, solves a whole-body withdrawal pose, and delivers it directly to the motors within 10 ms. Bottom-up reflex integration combines reflex-in-the-loop policy learning, reflex-transparent policy alignment, and reflex-informed policy adaptation to sustain the withdrawal, maintain balance, and complete longer-horizon recovery. We deploy ARC on a Unitree G1 and evaluate three categories of hazardous contact. Compared with an otherwise identical reflex delayed by 20 ms, ARC reduces the first force peak by 34.1% and reduces the contact impulse from 0.2632 N s to 0.1225 N s (a 2.15x ratio). We further demonstrate protected motion tracking, teleoperation, and vision-free obstacle traversal guided by foot tactile sensing. Project page: https://minusmodulo.github.io/reflexarc-website/

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong