REBot: Reflexive Evasion Robot Framework

• REBot is a quadrupedal control framework that uses reinforcement learning to achieve instantaneous dynamic obstacle avoidance with integrated avoidance and recovery policies.
• The system employs a finite-state machine with normal, avoidance, and recovery states to decouple rapid evasive maneuvers from post-evade stabilization, optimizing energy and stability.
• Empirical results in simulation and real-world experiments show that REBot outperforms baselines in short reaction-time regimes with higher avoidance and recovery success rates.

Reflexive Evasion Robot (REBot) is a control framework for quadrupedal robots that targets instantaneous dynamic obstacle avoidance (DOA) in regimes where conventional navigation-based trajectory replanning is too slow. Introduced as a reinforcement-learning-based system that integrates an avoidance policy and a recovery policy within a finite-state machine (FSM), REBot is designed for short reaction times, including cases below $1.5$ seconds, and emphasizes low-latency evasive maneuvers followed by rapid post-evasion stabilization (Xu et al., 8 Aug 2025). In the broader robotics literature, REBot also sits within a longer line of reflexive control ideas that includes predictive-and-reflexive manipulator planning (Miseikis et al., 2016), reactive whole-body obstacle avoidance in collaborative manipulation (Dastider et al., 2022), and the more general concept of reflexive input-output causality, in which actuation can itself become a new input for subsequent adaptation (Kayawake et al., 2024).

1. Conceptual scope and historical context

REBot addresses a specific failure mode of dynamic obstacle avoidance: existing approaches typically rely on navigation-based trajectory replanning, which assumes sufficient reaction time and can fail when obstacles approach rapidly (Xu et al., 8 Aug 2025). The framework therefore formalizes reflexive evasion as a control problem for legged robots operating under short reaction times, rather than as a variant of ordinary local planning.

This formulation has precedents. An earlier manipulator system combined multi-3D-camera workspace mapping with predictive and reflexive robot manipulator trajectory estimation, where reflexive behavior overrode predictive planning when an obstacle got unexpectedly close to the robot (Miseikis et al., 2016). In dual-arm collaborative robotics, SERA defined a reactive whole-body obstacle avoidance methodology that used proximity sensing, topological manifold learning, and graph-based replanning to reroute motion in real time when dynamic interference occurred (Dastider et al., 2022). At a more abstract level, reflexive robotics has also been framed as reflexive input-output causality, in which environmental or internal stimuli trigger immediate responses and those responses can be sequentially re-used as input to drive further adaptations (Kayawake et al., 2024).

Within this lineage, REBot is distinguished by its focus on quadrupedal locomotion, instantaneous evasive motion, and explicit separation between avoidance and recovery as coordinated learned behaviors (Xu et al., 8 Aug 2025).

2. Problem setting and state-transition logic

The REBot operating regime is defined by dynamic obstacles and extremely constrained reaction time. The framework is intended for scenarios in which quadrupedal robots require reflexive evasion capabilities to perform instantaneous, low-latency maneuvers, especially when $T_{\text{react}} < 1.5\text{s}$ (Xu et al., 8 Aug 2025).

Its control logic is organized as a three-state FSM:

• Normal State: the robot performs its baseline locomotion or primary task, typically under a PD controller.
• Avoidance State: triggered when an approaching obstacle is detected using the condition $\langle v^O_t,\; p^R_t - p^O_t \rangle > 0$.
• Recovery State: entered when instability is detected after evasion.

The instability conditions monitored for the recovery transition are stated in terms of body orientation, joint velocity, and base height: $$\|\theta^R_t\| > \theta^R_{\text{th}}, \qquad \|\dot q^R_t\| > \dot q^R_{\text{th}}, \qquad h^R_t < h^R_{\text{th}}.$$ The avoidance policy is trained to execute immediate evasive maneuvers, while the recovery policy is trained to regain stable posture and return the system to normal operation (Xu et al., 8 Aug 2025).

A common misconception is to treat REBot as merely a reactive fallback for a planner. The published formulation is narrower and more specific: it is a reflexive controller for instantaneous DOA, not a general-purpose planner. This distinction matters because later work argues that purely reactive decisions can sacrifice goal progress and passability, motivating hierarchical combinations of planning and reflexive evasion rather than replacement of one by the other (Xu et al., 7 Feb 2026).

3. Control architecture and policy representation

REBot integrates two learned policies inside the FSM: an RL-based avoidance policy and an RL-based recovery policy (Xu et al., 8 Aug 2025). Both are implemented as independent actor-critic neural networks, specifically MLPs, and are optimized with PPO.

The architecture is modular in a strong sense. The avoidance policy is responsible for instantaneous obstacle evasion under short reaction times, while the recovery policy focuses on rapid stabilization after aggressive maneuvers. The FSM then provides the arbitration logic needed to switch automatically between normal operation, avoidance, and recovery (Xu et al., 8 Aug 2025).

This separation is technically significant because evasive locomotion and stabilization impose different control objectives. The qualitative behaviors reported for REBot include jumping away from frontal or lateral threats and crouching for overhead threats, indicating that the learned policy space is not restricted to a single escape primitive (Xu et al., 8 Aug 2025). The paper also notes a mechanical design bias, specifically that it is easier to avoid frontal obstacles given rear leg power, which anchors the learned repertoire in the morphology of the platform rather than in purely kinematic command design (Xu et al., 8 Aug 2025).

The general idea of decoupling task behavior from a high-frequency safety layer also appears in related safe-RL work. Bresa separates a low-frequency task solver from a higher-frequency safety critic that can interrupt unsafe low-level execution in real time, explicitly positioning reflexive intervention at the control layer rather than only at the decision layer (Zhang et al., 27 Mar 2025). This suggests a broader design pattern in reflexive robotics: high-level intent and low-level safety are often learned or executed on different time scales.

4. Curriculum learning, reward shaping, and robustness mechanisms

REBot trains the avoidance policy with a two-stage curriculum (Xu et al., 8 Aug 2025). In Stage 1, a static obstacle appears near the robot, forcing reaction to abrupt presence and supporting learning of basic evasive maneuvers. In Stage 2, a moving obstacle with randomized speed and direction approaches, requiring adaptive real-time avoidance. The paper states that this progressive exposure promotes transfer and generalization.

The reward design is explicitly composite: $$r = r_{\text{avoidance}} + r_{\text{regularization}} + r_{\text{adaptive}}.$$

The avoidance reward contains a distance term and a collision term: $$r_{\text{distance}} = -\exp\left(-(d(p^O_t,\mathcal{B}^R)-r^O)\right),$$

$$r_{\text{collision}} = \mathbf{1}(c=0) - \mathbf{1}(c=1).$$

These encourage clearance and penalize contact (Xu et al., 8 Aug 2025).

The regularization reward encodes stability and efficiency. The gait-stability term rewards diagonal-leg synchronization for trot-like motions: $$r_{\text{walk}}= \frac{1}{2}\left( \mathbf{1}(c_{\text{FL}}=c_{\text{RR}}) + \mathbf{1}(c_{\text{FR}}=c_{\text{RL}}) \right).$$ The energy term penalizes power usage: $$r_{\text{energy}} = -\sum_i \left|\tau_t^{R,i}\cdot \dot q_t^{R,i}\right|,$$ and the contact-stability term penalizes high variance in foot contact forces: $$r_{\text{contact}} = - \sum_i \left(f_t^{R,i,z}-f_{t-1}^{R,i,z}\right)^2.$$ These terms are described as promoting stable, efficient, and physically plausible gaits (Xu et al., 8 Aug 2025).

The adaptive reward is aimed at robustness and behavioral diversity. It includes a diversity term,

$T_{\text{react}} < 1.5\text{s}$0

a threat-adaptation term,

$T_{\text{react}} < 1.5\text{s}$1

and a directional-efficiency term,

$T_{\text{react}} < 1.5\text{s}$2

The recovery reward is described as being designed to maximize upright orientation restoration, minimize joint velocity, penalize large posture deviations, and smooth control transitions (Xu et al., 8 Aug 2025).

Robustness is further supported by domain randomization over simulation parameters including joint state noise, base mass, friction, and obstacle velocity, size, and position (Xu et al., 8 Aug 2025). A plausible implication is that REBot’s reported sim-to-real transfer depends as much on this randomized training distribution as on the specific policy architecture.

5. Empirical evaluation, metrics, and observed limitations

REBot is evaluated in Isaac Gym simulation and in real-world experiments on a Unitree Go2 (Xu et al., 8 Aug 2025). The reported metrics are Avoidance Success Rate (ASR), Recovery Success Rate (RSR), Maximum Joint Power (MJP), Avoidance Moving Distance (AMD), and Gait Diversity Index (GDI).

In simulation, the main reported results are split by reaction-time regime:

Reaction-time regime	REBot	Baselines
$T_{\text{react}} < 1.5\text{s}$3 s, ASR	0.65	ABS 0.11, RRL 0.09
$T_{\text{react}} < 1.5\text{s}$4 s, RSR	0.59	ABS 0.06, RRL 0.05
$T_{\text{react}} < 1.5\text{s}$5 s, MJP	0.49	ABS 0.52, RRL 0.51
$T_{\text{react}} < 1.5\text{s}$6 s, AMD	0.47	ABS 0.80, RRL 0.86
$T_{\text{react}} < 1.5\text{s}$7 s, ASR	0.81	ABS 0.51, RRL 0.41
$T_{\text{react}} < 1.5\text{s}$8 s, RSR	0.74	ABS 0.42, RRL 0.32
$T_{\text{react}} < 1.5\text{s}$9 s, MJP	0.34	ABS 0.40, RRL 0.45
$\langle v^O_t,\; p^R_t - p^O_t \rangle > 0$0 s, AMD	0.26	ABS 0.60, RRL 0.70

The paper summarizes these results as marked improvement over both ABS and RRL baselines in all reaction-time regimes except the extreme lower bound $\langle v^O_t,\; p^R_t - p^O_t \rangle > 0$1 s, with especially strong gains in the short reaction-time reflex regime (Xu et al., 8 Aug 2025). Lower MJP and AMD are interpreted in the paper as indicators of more efficient, controlled evasions.

Ablation results identify several dependencies. Removing the recovery module produces a 20% drop in recovery success in the reflex regime $\langle v^O_t,\; p^R_t - p^O_t \rangle > 0$2. Removing curriculum learning causes a 5% decrease in avoidance and recovery rates. Removing the adaptive reward leads to a 40% decrease in gait diversity and lower ASR/RSR (Xu et al., 8 Aug 2025). These ablations directly support the view that REBot is not only an avoidance policy, but a coupled avoidance-plus-recovery system.

Real-world trials were conducted with three dynamic threats: stick pokes, thrown balls, and human kicks. The reported results are ASR $\langle v^O_t,\; p^R_t - p^O_t \rangle > 0$3 and RSR $\langle v^O_t,\; p^R_t - p^O_t \rangle > 0$4 (Xu et al., 8 Aug 2025). The paper attributes the performance gap relative to simulation to unmodelled motor characteristics, control latency, and surface friction. This is an important limitation: the reflexive control paradigm transfers meaningfully to hardware, but instantaneous evasion remains sensitive to actuation and contact-model mismatch.

6. Extensions, related systems, and evolving interpretations

Subsequent and adjacent works place REBot inside a broader systems stack. APREBot introduces an active perception system for Reflexive Evasion Robot that combines LiDAR-based omnidirectional scanning with camera-based active focusing and a hierarchical threat-aware controller; its reflexive evasion stage uses a learned deep-RL reflex policy adapted from REBot (Xu et al., 29 Sep 2025). On a Unitree Go2 with a Livox MID-360 LiDAR and Intel RealSense D435 RGB-D camera, APREBot reports ASR $\langle v^O_t,\; p^R_t - p^O_t \rangle > 0$5, minimum clearance $\langle v^O_t,\; p^R_t - p^O_t \rangle > 0$6 m, TNL $\langle v^O_t,\; p^R_t - p^O_t \rangle > 0$7, and energy $\langle v^O_t,\; p^R_t - p^O_t \rangle > 0$8 J, outperforming Raycast, FAPP, and Omni baselines on the metrics reported in its Table 1 (Xu et al., 29 Sep 2025). This suggests that reflexive evasion alone is not the full problem; onboard perception and threat prioritization are equally central when obstacles can emerge from any direction.

UEREBot generalizes the idea further by separating slow planning from instantaneous reflexive evasion and coordinating them through a threat-aware handoff and a control barrier function shield (Xu et al., 7 Feb 2026). Its reported motivation is that a single system cannot fully satisfy long-horizon goal progress, passability, and collision avoidance against high-speed dynamic obstacles simultaneously. In its main results, UEREBot reports GCR $\langle v^O_t,\; p^R_t - p^O_t \rangle > 0$9, ASR $$\|\theta^R_t\| > \theta^R_{\text{th}}, \qquad \|\dot q^R_t\| > \dot q^R_{\text{th}}, \qquad h^R_t < h^R_{\text{th}}.$$0, TSR $$\|\theta^R_t\| > \theta^R_{\text{th}}, \qquad \|\dot q^R_t\| > \dot q^R_{\text{th}}, \qquad h^R_t < h^R_{\text{th}}.$$1, PE $$\|\theta^R_t\| > \theta^R_{\text{th}}, \qquad \|\dot q^R_t\| > \dot q^R_{\text{th}}, \qquad h^R_t < h^R_{\text{th}}.$$2, $$\|\theta^R_t\| > \theta^R_{\text{th}}, \qquad \|\dot q^R_t\| > \dot q^R_{\text{th}}, \qquad h^R_t < h^R_{\text{th}}.$$3, and LC $$\|\theta^R_t\| > \theta^R_{\text{th}}, \qquad \|\dot q^R_t\| > \dot q^R_{\text{th}}, \qquad h^R_t < h^R_{\text{th}}.$$4, exceeding representative baselines on ASR, TSR, PE, and minimum clearance (Xu et al., 7 Feb 2026). A plausible implication is that REBot functions effectively as the instantaneous layer of a larger autonomy hierarchy rather than as a standalone solution to locomotion in unstructured environments.

The reflexive-evasion concept has also diversified across embodiments. AWARE addresses high-dynamic reflexive obstacle evasion in wheeled-legged robots through hierarchical reinforcement learning, explicit mode selection, and dual low-level experts, yielding emergent behaviors such as forward lunge and lateral dodge; it reports simulation ASR of 97.3% for $$\|\theta^R_t\| > \theta^R_{\text{th}}, \qquad \|\dot q^R_t\| > \dot q^R_{\text{th}}, \qquad h^R_t < h^R_{\text{th}}.$$5 s reaction time and real-world ASR of 59% on the M20 platform (Zhao et al., 26 Apr 2026). Earlier manipulator systems achieved reflexive backtracking or vector evasion when obstacles approached unexpectedly (Miseikis et al., 2016), and SERA demonstrated reactive whole-body avoidance for dual-arm systems using local sensing and manifold-based replanning, with runtime and success-ratio improvements over Dynamic RRT* and MpNet baselines (Dastider et al., 2022). These works indicate that “reflexive evasion robot” is not limited to a single morphology.

At the level of mechanism, REBot should also be distinguished from physically embodied notions of reflexive actuation. The reflexive input-output causality framework includes contact-sensitive reflexive arms, pneumatic and bellows systems, intra-effectors, and chained reflexes in which outputs become new inputs without centralized computation (Kayawake et al., 2024). REBot, by contrast, is a software-centric quadrupedal control framework built around PPO-trained policies and FSM arbitration (Xu et al., 8 Aug 2025). The two are compatible but not equivalent: one emphasizes embodied causal coupling, the other learned low-latency locomotor response.

Taken together, the literature positions REBot as a specific and influential formulation of reflexive evasion for quadrupeds: an RL-driven FSM that couples instantaneous avoidance with stabilizing recovery, performs best in short reaction-time regimes relative to the baselines tested, and increasingly serves as a module inside larger perception-and-planning architectures rather than as a complete autonomy stack on its own (Xu et al., 8 Aug 2025).