Adaptive Degeneracy Optimization Agent (DOA)

• The paper introduces a deep reinforcement learning agent that adaptively compensates for degeneracy in particle filter-based 2D-SLAM by learning a continuous degeneracy factor.
• It employs an actor–critic architecture with PPO and multi-head attention to fuse sensor data, optimizing pose estimation through adaptive interpolation between motion and observation models.
• Experimental results show up to an 80% reduction in absolute translational error and a 90% degeneracy detection success rate in both simulated and real indoor environments.

Searching arXiv for the specified DOA paper and closely related SLAM degeneracy work. Adaptive Degeneracy Optimization Agent (DOA) is a deep reinforcement learning method for mitigating degeneracy in particle filter-based 2D-SLAM, particularly in indoor environments such as long straight corridors where scan matching becomes unreliable. The method, introduced in “DOA: A Degeneracy Optimization Agent with Adaptive Pose Compensation Capability based on Deep Reinforcement Learning” (Li et al., 26 Jul 2025), uses Proximal Policy Optimization (PPO) to train an agent that estimates a scalar degeneracy factor and then adaptively reweights the relative contribution of the observation distribution and the motion model distribution during pose optimization. In the reported formulation, DOA is designed to address three challenges associated with traditional supervised learning frameworks—data acquisition bottlenecks in degenerate dataset, inherent quality deterioration of training samples, and ambiguity in annotation protocol design—while also providing transfer learning-based generalization across environments (Li et al., 26 Jul 2025).

1. Problem setting and formalization

Particle filter-based 2D-SLAM is described as widely used in indoor localization tasks due to its efficiency, but long straight corridors and other symmetric indoor environments can induce severe degeneracy problems (Li et al., 26 Jul 2025). Within the reported SLAM back-end, GMapping maintains two particle distributions at each time step $t$. The first is the motion-model distribution $p_u$, whose particles are sampled from

$p(x_t|x_{t-1},u_t)=N(f(x_{t-1},u_t),\Sigma_{u_m}).$

The second is the observation distribution $p_z$, obtained by scan matching (Li et al., 26 Jul 2025).

The paper places this within the standard SLAM posterior decomposition

$$p(x_t,M|z_t,u_t)=p(M|x_t,z_{1:t})\cdot p(x_t|z_t,u_t),$$

and specifies the observation model as

$$p(z_t|x_t,M)\propto \exp\left[-\frac{1}{2}\sum_i \|z_t^{(i)}-h(x_t,M)\|^2_{\Sigma_{z_t}^{-1}}\right].$$

Degeneracy is characterized through rank-deficiency in the Fisher information structure,

$$I(x_t)=I_{z_m}+I_{u_m}, \qquad I_{z_m}\approx \mathrm{diag}(\lambda_x\to 0,\lambda_y,\lambda_\theta),$$

which captures the loss of observability along a degenerate direction (Li et al., 26 Jul 2025).

A central quantity in the method is the “degree of degeneracy,” interpreted as the relative unreliability of $p_z$ versus $p_u$. An analytic proxy is the Mahalanobis distance

$D_M=\sqrt{(g_u-g_z)^T\Sigma_{x_t}^{-1}(g_u-g_z)},$

where $p_u$0 and $p_u$1 are the centroids of the two particle sets and $p_u$2 is the combined covariance. Large $p_u$3 indicates large drift of $p_u$4 along a degenerate direction (Li et al., 26 Jul 2025).

2. Agent architecture and state representation

DOA operates directly on particle-level state information from the SLAM back-end. At each step, the coordinates of the particles from the motion-model distribution and the observation distribution are concatenated into a flattened vector $p_u$5. In practice, $p_u$6, yielding $p_u$7 (Li et al., 26 Jul 2025).

Before entering the learned network, this state is transformed by a fixed “special weight matrix” $p_u$8. The reported purpose of this matrix is to encourage the network to learn global distribution patterns (Li et al., 26 Jul 2025). The policy network then consists of four linear+ReLU layers, with the first transformation given by the fixed matrix stage $p_u$9, followed by $p(x_t|x_{t-1},u_t)=N(f(x_{t-1},u_t),\Sigma_{u_m}).$0, $p(x_t|x_{t-1},u_t)=N(f(x_{t-1},u_t),\Sigma_{u_m}).$1, and optionally a fourth $p(x_t|x_{t-1},u_t)=N(f(x_{t-1},u_t),\Sigma_{u_m}).$2 layer for richer features. The neck contains two linear+ReLU layers $p(x_t|x_{t-1},u_t)=N(f(x_{t-1},u_t),\Sigma_{u_m}).$3 and then a 3-head multi-head attention module, with each head 64-dimensional, to capture complex inter-particle relationships and subtle degeneracy cues, including Mahalanobis spread between $p(x_t|x_{t-1},u_t)=N(f(x_{t-1},u_t),\Sigma_{u_m}).$4 and $p(x_t|x_{t-1},u_t)=N(f(x_{t-1},u_t),\Sigma_{u_m}).$5 (Li et al., 26 Jul 2025).

The actor head applies linear+ReLU $p(x_t|x_{t-1},u_t)=N(f(x_{t-1},u_t),\Sigma_{u_m}).$6 followed by linear+Sigmoid $p(x_t|x_{t-1},u_t)=N(f(x_{t-1},u_t),\Sigma_{u_m}).$7, outputting a degeneracy factor $p(x_t|x_{t-1},u_t)=N(f(x_{t-1},u_t),\Sigma_{u_m}).$8. The critic shares the same 256-dimensional backbone and neck; its value head uses three linear+ReLU layers $p(x_t|x_{t-1},u_t)=N(f(x_{t-1},u_t),\Sigma_{u_m}).$9 followed by a final linear layer to predict $p_z$0 (Li et al., 26 Jul 2025). The value function is given as

$p_z$1

The architecture is notable in that the output is not a discrete label but a continuous control variable. During training, the network learns to map raw particle coordinates and their divergence statistics to $p_z$2, where $p_z$3 denotes no degeneracy and $p_z$4 denotes high degeneracy (Li et al., 26 Jul 2025). This suggests that DOA treats degeneracy estimation and compensation as a coupled inference-and-control problem rather than a purely diagnostic task.

3. Degeneracy factor and adaptive pose compensation

The defining mechanism of DOA is the use of the learned degeneracy factor $p_z$5 as a reference weight for sensor contribution in pose optimization. Given $p_z$6, the back-end fuses the observation distribution and the motion-model distribution by shifting the centroid of $p_z$7 toward $p_z$8 along the line connecting them (Li et al., 26 Jul 2025). The fusion equations are

$p_z$9

and

$$p(x_t,M|z_t,u_t)=p(M|x_t,z_{1:t})\cdot p(x_t|z_t,u_t),$$0

For the mean alone, the paper gives the equivalent closed-form linear interpolation

$$p(x_t,M|z_t,u_t)=p(M|x_t,z_{1:t})\cdot p(x_t|z_t,u_t),$$1

The limiting behavior is explicit: in the degenerate limit $$p(x_t,M|z_t,u_t)=p(M|x_t,z_{1:t})\cdot p(x_t|z_t,u_t),$$2 along one axis, $$p(x_t,M|z_t,u_t)=p(M|x_t,z_{1:t})\cdot p(x_t|z_t,u_t),$$3 and $$p(x_t,M|z_t,u_t)=p(M|x_t,z_{1:t})\cdot p(x_t|z_t,u_t),$$4, corresponding to full trust in the motion model (Li et al., 26 Jul 2025). This mechanism operationalizes the statement that the observation distribution is shifted towards the motion model distribution, with the step size determined by a linear interpolation formula related to the degeneracy factor.

Within the reported pseudoalgorithm, the agent extracts particle coordinates from $$p(x_t,M|z_t,u_t)=p(M|x_t,z_{1:t})\cdot p(x_t|z_t,u_t),$$5 and $$p(x_t,M|z_t,u_t)=p(M|x_t,z_{1:t})\cdot p(x_t|z_t,u_t),$$6, forwards the resulting state through the actor to obtain $$p(x_t,M|z_t,u_t)=p(M|x_t,z_{1:t})\cdot p(x_t|z_t,u_t),$$7, applies the fusion to produce $$p(x_t,M|z_t,u_t)=p(M|x_t,z_{1:t})\cdot p(x_t|z_t,u_t),$$8, and then chooses the best pose among $$p(x_t,M|z_t,u_t)=p(M|x_t,z_{1:t})\cdot p(x_t|z_t,u_t),$$9 and $$p(z_t|x_t,M)\propto \exp\left[-\frac{1}{2}\sum_i \|z_t^{(i)}-h(x_t,M)\|^2_{\Sigma_{z_t}^{-1}}\right].$$0 by likelihood (Li et al., 26 Jul 2025). A plausible implication is that DOA is not merely suppressing scan matching under degeneracy; rather, it performs a continuous interpolation between observation-driven and motion-driven localization.

4. Reinforcement learning objective and training methodology

DOA is trained with PPO rather than supervised labels. The PPO objective is written as

$$p(z_t|x_t,M)\propto \exp\left[-\frac{1}{2}\sum_i \|z_t^{(i)}-h(x_t,M)\|^2_{\Sigma_{z_t}^{-1}}\right].$$1

with

$$p(z_t|x_t,M)\propto \exp\left[-\frac{1}{2}\sum_i \|z_t^{(i)}-h(x_t,M)\|^2_{\Sigma_{z_t}^{-1}}\right].$$2

The reward at each step is

$$p(z_t|x_t,M)\propto \exp\left[-\frac{1}{2}\sum_i \|z_t^{(i)}-h(x_t,M)\|^2_{\Sigma_{z_t}^{-1}}\right].$$3

Here, $$p(z_t|x_t,M)\propto \exp\left[-\frac{1}{2}\sum_i \|z_t^{(i)}-h(x_t,M)\|^2_{\Sigma_{z_t}^{-1}}\right].$$4 are the posterior variances of the optimized pose, encouraging contraction of uncertainty; $$p(z_t|x_t,M)\propto \exp\left[-\frac{1}{2}\sum_i \|z_t^{(i)}-h(x_t,M)\|^2_{\Sigma_{z_t}^{-1}}\right].$$5 is the scan-matching likelihood score of the best particle, encouraging fit to map; and

$$p(z_t|x_t,M)\propto \exp\left[-\frac{1}{2}\sum_i \|z_t^{(i)}-h(x_t,M)\|^2_{\Sigma_{z_t}^{-1}}\right].$$6

is the effective particle count after normalization, encouraging particle diversity, with

$$p(z_t|x_t,M)\propto \exp\left[-\frac{1}{2}\sum_i \|z_t^{(i)}-h(x_t,M)\|^2_{\Sigma_{z_t}^{-1}}\right].$$7

The final term $$p(z_t|x_t,M)\propto \exp\left[-\frac{1}{2}\sum_i \|z_t^{(i)}-h(x_t,M)\|^2_{\Sigma_{z_t}^{-1}}\right].$$8 penalizes abrupt changes in the degeneracy factor and stabilizes the policy (Li et al., 26 Jul 2025). The empirical weights are $$p(z_t|x_t,M)\propto \exp\left[-\frac{1}{2}\sum_i \|z_t^{(i)}-h(x_t,M)\|^2_{\Sigma_{z_t}^{-1}}\right].$$9, $$I(x_t)=I_{z_m}+I_{u_m}, \qquad I_{z_m}\approx \mathrm{diag}(\lambda_x\to 0,\lambda_y,\lambda_\theta),$$0, $$I(x_t)=I_{z_m}+I_{u_m}, \qquad I_{z_m}\approx \mathrm{diag}(\lambda_x\to 0,\lambda_y,\lambda_\theta),$$1, and $$I(x_t)=I_{z_m}+I_{u_m}, \qquad I_{z_m}\approx \mathrm{diag}(\lambda_x\to 0,\lambda_y,\lambda_\theta),$$2.

The reported PPO hyperparameters are learning rate $$I(x_t)=I_{z_m}+I_{u_m}, \qquad I_{z_m}\approx \mathrm{diag}(\lambda_x\to 0,\lambda_y,\lambda_\theta),$$3, discount $$I(x_t)=I_{z_m}+I_{u_m}, \qquad I_{z_m}\approx \mathrm{diag}(\lambda_x\to 0,\lambda_y,\lambda_\theta),$$4, clip $$I(x_t)=I_{z_m}+I_{u_m}, \qquad I_{z_m}\approx \mathrm{diag}(\lambda_x\to 0,\lambda_y,\lambda_\theta),$$5, epochs per update $$I(x_t)=I_{z_m}+I_{u_m}, \qquad I_{z_m}\approx \mathrm{diag}(\lambda_x\to 0,\lambda_y,\lambda_\theta),$$6, timesteps per rollout $$I(x_t)=I_{z_m}+I_{u_m}, \qquad I_{z_m}\approx \mathrm{diag}(\lambda_x\to 0,\lambda_y,\lambda_\theta),$$7, and total episodes $$I(x_t)=I_{z_m}+I_{u_m}, \qquad I_{z_m}\approx \mathrm{diag}(\lambda_x\to 0,\lambda_y,\lambda_\theta),$$8 (Li et al., 26 Jul 2025). Training uses scene sampling with non-degenerate environments for initial training and degenerate environments, specifically long corridors, for fine-tuning. A transfer learning module is then applied: after convergence in non-degenerate scenes, the backbone and neck are frozen, and the actor head and critic are fine-tuned in degenerate scenes. The stated purpose is to preserve general feature extraction while adapting to feature-poor corridors and to address the inefficiency of training in degenerate environments (Li et al., 26 Jul 2025).

5. Experimental evaluation and ablation results

The evaluation covers six environments: four simulated environments, denoted S1–S4, and two real environments, denoted R1–R2 (Li et al., 26 Jul 2025). For degeneracy detection, the threshold $$I(x_t)=I_{z_m}+I_{u_m}, \qquad I_{z_m}\approx \mathrm{diag}(\lambda_x\to 0,\lambda_y,\lambda_\theta),$$9 is considered degenerate. Under this criterion, DOA achieves approximately $p_z$0 success rate, compared with Switch-SLAM at approximately $p_z$1 and Li et al. at approximately $p_z$2. The reported inference time is approximately $p_z$3, which is less than the SLAM update time of $p_z$4 (Li et al., 26 Jul 2025).

Ablation studies compare PPO with and without critic, with and without transfer learning, and alternative algorithms, specifically DDPG and SAC, under the same network architecture (Li et al., 26 Jul 2025). Removing the critic or transfer learning reduces success by approximately $p_z$5–$p_z$6. This supports the stated role of both the actor–critic formulation and the transfer learning stage.

For degeneracy optimization in simulation, the absolute translational error (ATE) of GMapping+DOA is reduced by $p_z$7 in S1 and by more than $p_z$8 in deeper corridors S2–S4, outperforming prior anti-degeneracy methods (Li et al., 26 Jul 2025). In real robot experiments R1–R2, the reported outcome is qualitative map correctness, with corridor lengths and landmark alignments matching ground truth. The paper also states that resample frequency is lowered by DOA, indicating more stable localization (Li et al., 26 Jul 2025).

Evaluation aspect	Reported result
Degeneracy detection environments	6 environments: S1–S4, R1–R2
Degeneracy threshold	$p_z$9 considered degenerate
Detection success rate	DOA $p_u$0
Baseline detection rates	Switch-SLAM $p_u$1; Li et al. $p_u$2
Inference time	$p_u$3
SLAM update time	$p_u$4
ATE reduction in S1	$p_u$5
ATE reduction in S2–S4	$p_u$6

These results are presented as evidence that DOA improves both degeneracy detection and downstream localization quality across simulated and real environments (Li et al., 26 Jul 2025).

6. Algorithmic workflow, interpretation, and related issues

The reported algorithmic summary is iterative and tightly integrated with GMapping. The process begins by initializing actor–critic networks and the GMapping particle filter. For each episode, the robot pose is reset in a simulated or real scene. At each SLAM update, the system extracts 30-particle coordinates from the motion model $p_u$7 and scan-matching distribution $p_u$8, forms the state $p_u$9, forwards it through the actor to obtain $D_M=\sqrt{(g_u-g_z)^T\Sigma_{x_t}^{-1}(g_u-g_z)},$0, applies the fusion to shift $D_M=\sqrt{(g_u-g_z)^T\Sigma_{x_t}^{-1}(g_u-g_z)},$1 toward $D_M=\sqrt{(g_u-g_z)^T\Sigma_{x_t}^{-1}(g_u-g_z)},$2 by factor $D_M=\sqrt{(g_u-g_z)^T\Sigma_{x_t}^{-1}(g_u-g_z)},$3, chooses the best pose among $D_M=\sqrt{(g_u-g_z)^T\Sigma_{x_t}^{-1}(g_u-g_z)},$4 and $D_M=\sqrt{(g_u-g_z)^T\Sigma_{x_t}^{-1}(g_u-g_z)},$5 by likelihood, computes the reward using covariance, likelihood, $D_M=\sqrt{(g_u-g_z)^T\Sigma_{x_t}^{-1}(g_u-g_z)},$6, and $D_M=\sqrt{(g_u-g_z)^T\Sigma_{x_t}^{-1}(g_u-g_z)},$7, stores the transition tuple, and advances the SLAM map. After $D_M=\sqrt{(g_u-g_z)^T\Sigma_{x_t}^{-1}(g_u-g_z)},$8, PPO updates are performed for $D_M=\sqrt{(g_u-g_z)^T\Sigma_{x_t}^{-1}(g_u-g_z)},$9, and in the transfer-learning phase the backbone and neck are frozen before fine-tuning (Li et al., 26 Jul 2025).

Several interpretive points follow from this design. First, the approach does not rely on manually tuned thresholds for compensation; rather, it learns compensation behavior through reinforcement signals. Second, the scalar output $p_u$00 serves both as a degeneracy detector and as a control variable for adaptive sensor fusion. Third, the paper explicitly positions reinforcement learning as a response to three limitations of supervised learning frameworks: data acquisition bottlenecks in degenerate dataset, inherent quality deterioration of training samples, and ambiguity in annotation protocol design (Li et al., 26 Jul 2025).

A common misconception would be to treat DOA as a replacement for the particle filter itself. The reported formulation instead operates as an optimization agent on top of the SLAM back-end, using the existing motion-model and scan-matching particle sets as state and modifying the pose estimate through adaptive fusion. Another possible misconception is that degeneracy is handled by a hard switch from observation to odometry. The use of $p_u$01 and the interpolation equations shows that the method is continuous rather than binary (Li et al., 26 Jul 2025).

In broader methodological terms, DOA sits at the intersection of particle filter SLAM, anti-degeneracy pose optimization, and actor–critic reinforcement learning. The paper’s emphasis on transfer learning across non-degenerate and degenerate scenes suggests a research direction in which feature extraction for particle distribution geometry is learned in rich environments and then specialized for feature-poor corridors. A plausible implication is that the method’s main contribution lies not only in degeneracy detection accuracy but also in embedding that detection into the pose update itself through a differentiable, continuous compensation mechanism (Li et al., 26 Jul 2025).