AlphaAdj: Adaptive CBF Navigation

• AlphaAdj is a vision-to-control navigation framework that adapts the control barrier function parameter using semantic risk from egocentric RGB images.
• It employs a dynamic cap combining geometric clearance and speed factors to balance safety and efficiency in cluttered, dynamic environments.
• The system uses a staleness-gated fusion policy to ensure robust, collision-free navigation despite asynchronous VLM inference and latency issues.

Searching arXiv for the AlphaAdj paper and closely related safe-navigation/CBF work. AlphaAdj is a vision-to-control navigation framework that adapts the conservativeness of a control barrier function (CBF) safety filter online using semantic risk inferred from egocentric RGB images by a vision-LLM (VLM). In the method introduced as “AlphaAdj,” the adapted quantity is the CBF parameter $\alpha$, which appears in the CBF safety inequality and governs how strongly the barrier constraint is enforced. The framework targets single-robot safe navigation in cluttered, possibly dynamic, unstructured environments, and combines a bounded risk-to-$\alpha$ mapping, a geometric and speed-aware dynamic cap, and a staleness-gated fusion policy to handle asynchronous VLM inference and latency while preserving collision-free navigation and improving efficiency relative to fixed CBF settings (Chen et al., 22 Mar 2026).

1. Problem setting and core definition

AlphaAdj is formulated for single-robot navigation with egocentric RGB observations in environments that may contain static obstacles, moving obstacles, narrow passages, clutter, or partially occluded hazards. The robot must reach a goal while avoiding collisions. The method is motivated by the limitation of fixed-parameter CBF safety filters: if $\alpha$ is small, the filter is conservative and tends to intervene early, which is safe but can induce long detours, hesitations, and failures to make progress; if $\alpha$ is large, the filter allows more aggressive motion and can be efficient in benign scenes but may become too permissive near hazards (Chen et al., 22 Mar 2026).

The central design choice is to adapt the scalar CBF parameter $\alpha(t)\in[\alpha_{\min},\alpha_{\max}]$ online. Smaller $\alpha$ yields more conservative behavior near obstacles, while larger $\alpha$ yields more aggressive behavior. AlphaAdj therefore does not replace the CBF safety filter; it changes the aggressiveness of the existing safety mechanism by modulating the parameter that enters the barrier inequality.

At a high level, the pipeline begins with the robot’s egocentric RGB image $o(t)$. A VLM processes this image asynchronously and outputs a scalar collision-risk estimate $r\in[0,1]$, where low values mean benign scenes and high values mean imminent hazard. That risk estimate is mapped to a candidate $\alpha_{\text{vlm}}$. In parallel, the robot computes geometric safety quantities such as the minimum barrier margin to the nearest obstacle and its current speed, which define a dynamic cap $\alpha$0. A staleness test determines whether the most recent VLM output is still fresh enough to trust. The final parameter $\alpha$1 is then passed into the CBF safety filter, which sits on top of a nominal goal-seeking controller (Chen et al., 22 Mar 2026).

2. Control-theoretic formulation

The robot is modeled as a discrete-time control-affine system: $\alpha$2 where $\alpha$3 includes planar position $\alpha$4 and heading $\alpha$5, and $\alpha$6 is a desired planar velocity vector in the world frame. This desired planar velocity is later converted to differential-drive commands $\alpha$7 (Chen et al., 22 Mar 2026).

Safety is encoded through a control barrier function $\alpha$8, with safe set

$\alpha$9

The CBF condition is

$\alpha$0

and under the control-affine dynamics this becomes

$\alpha$1

This is the inequality that AlphaAdj modulates. The interpretation is explicit: smaller $\alpha$2 makes the right-hand side less negative and forces $\alpha$3 to be larger, producing more conservative behavior; larger $\alpha$4 is more permissive (Chen et al., 22 Mar 2026).

For obstacle avoidance, the implementation uses a per-obstacle squared-distance barrier. If $\alpha$5 is the robot position, $\alpha$6 the center of circular obstacle $\alpha$7, $\alpha$8 the obstacle radius, $\alpha$9 the robot’s effective radius, and $\alpha$0 an optional padding, then

$\alpha$1

and the barrier for obstacle $\alpha$2 is

$\alpha$3

Positive $\alpha$4 means the robot is outside the padded obstacle, zero is the safety boundary, and negative means violation or collision (Chen et al., 22 Mar 2026).

The nominal controller proposes a goal-directed velocity $\alpha$5, limited by $\alpha$6. The safety filter then solves the quadratic program

$\alpha$7

subject to the CBF constraints and

$\alpha$8

A plausible implication is that AlphaAdj should be understood as a parameter-adaptive layer inserted into an otherwise standard CBF-QP pipeline, rather than as a new QP formulation.

3. Semantic risk estimation and risk-to-$\alpha$9 mapping

The VLM output is a bounded scalar risk estimate $\alpha(t)\in[\alpha_{\min},\alpha_{\max}]$0. The paper explicitly states that this score should be interpreted as “a practical proxy for scene difficulty or hazard likelihood in the near term” and “a hazard score for the robot’s current motion context,” rather than as a formal probabilistic risk metric (Chen et al., 22 Mar 2026).

The server-side prompt is specified as:

You are a risk estimator for mobile robot navigation. Given this scene image, output ONLY a single-line JSON of the form {"risk": <number between 0 and 1>}. No prose, no markdown---just JSON.

The controller augments this with online context:

You are a safety estimator for a mobile robot. Given the current RGB view from a camera mounted on the robot, estimate current collision risk as a number in [0,1]. Lower values meaning lower imminent collision risk, and higher values meaning imminent collision risk. Return strict JSON only with keys: risk (0..1 float). Context: $\alpha(t)\in[\alpha_{\min},\alpha_{\max}]$1 and current speed $\alpha(t)\in[\alpha_{\min},\alpha_{\max}]$2 m/s.

The risk value is clipped or rescaled to $\alpha(t)\in[\alpha_{\min},\alpha_{\max}]$3 if necessary. The paper also notes that few-shot risk exemplars are avoided to prevent anchoring the model to arbitrary numeric targets (Chen et al., 22 Mar 2026).

The candidate CBF parameter induced by semantics is defined by the monotone decreasing map

$\alpha(t)\in[\alpha_{\min},\alpha_{\max}]$4

with $\alpha(t)\in[\alpha_{\min},\alpha_{\max}]$5 a sensitivity parameter. If $\alpha(t)\in[\alpha_{\min},\alpha_{\max}]$6, then $\alpha(t)\in[\alpha_{\min},\alpha_{\max}]$7; if $\alpha(t)\in[\alpha_{\min},\alpha_{\max}]$8, then $\alpha(t)\in[\alpha_{\min},\alpha_{\max}]$9. Low perceived risk therefore leads to permissive behavior, while high perceived risk lowers $\alpha$0 and makes the CBF intervene earlier and more strongly (Chen et al., 22 Mar 2026).

The authors motivate this power-law map by boundedness, monotonicity, smoothness, and tunability through a single exponent $\alpha$1. This suggests that AlphaAdj is intentionally designed as a low-dimensional semantic interface to the safety filter: the VLM does not directly output actions, trajectories, or QP constraints, but only a bounded scalar that modulates conservativeness.

4. Dynamic cap and staleness-gated fusion

A major practical issue is that VLM inference is asynchronous and slow relative to the control loop. AlphaAdj therefore supplements the semantic map with a purely geometric and kinematic permissiveness cap and a freshness check for the latest VLM result (Chen et al., 22 Mar 2026).

The dynamic cap begins from the clearance margin $\alpha$2, defined as the distance between the robot and its closest obstacle. Large $\alpha$3 means more clearance; $\alpha$4 indicates collision with the nearest obstacle. Clearance is normalized as

$\alpha$5

where $\alpha$6 is a clearance scale. The method then interpolates between a conservative near value $\alpha$7 and a permissive far value $\alpha$8: $\alpha$9 with $\alpha$0 in the paper, so that the cap remains conservative until the robot is clearly far from obstacles (Chen et al., 22 Mar 2026).

Distance alone is not deemed sufficient, so the cap is tightened according to speed. With current forward speed $\alpha$1 and maximum allowed forward speed $\alpha$2,

$\alpha$3

where $\alpha$4. The final cap is

$\alpha$5

with

$\alpha$6

This cap does not depend on the VLM and acts as a safety envelope limiting how permissive $\alpha$7 may become (Chen et al., 22 Mar 2026).

Freshness is encoded by the staleness indicator

$\alpha$8

where $\alpha$9 is the time the latest valid VLM estimate was received. When the VLM estimate is fresh, AlphaAdj uses

$o(t)$0

When the estimate is stale, the VLM output is ignored and the policy uses

$o(t)$1

The soft and hard caps share the same equations for the cap, but the near-obstacle floor satisfies

$o(t)$2

Thus stale predictions force a more conservative minimum $o(t)$3 near obstacles (Chen et al., 22 Mar 2026).

This staleness-gated fusion policy is one of the method’s defining features. The semantic signal is never trusted unconditionally: it is both upper-bounded by geometry and disabled when too old.

5. System architecture and implementation

AlphaAdj operates as a two-rate control architecture. The low-level controller runs at $o(t)$4 Hz, corresponding to every $o(t)$5 seconds. Every $o(t)$6 control steps, the system captures an RGB frame and queries the VLM asynchronously over HTTP, which corresponds to $o(t)$7 Hz. The average end-to-end round-trip latency is reported as $o(t)$8 ms (Chen et al., 22 Mar 2026).

The control loop continues uninterrupted while waiting for VLM responses. The most recent valid $o(t)$9 is held until another response arrives. To reduce latency, the implementation uses strict JSON-only outputs, short scripted responses to reduce token generation, local hosting rather than cloud inference, and 4-bit quantization of model weights. These changes reduce average end-to-end latency from $r\in[0,1]$0 ms to $r\in[0,1]$1 ms, a $r\in[0,1]$2 improvement (Chen et al., 22 Mar 2026).

The simulation stack uses IsaacSim with a Carter differential-drive robot and an egocentric RGB camera. The system runs IsaacSim on one GPU and a local FastAPI-served LLaVA-1.5 7B VLM on another. The paper states that LLaVA-1.5 7B was chosen because it offered a latency-quality tradeoff better than larger variants such as 13B (Chen et al., 22 Mar 2026).

The robot starts at rest and is bounded to a maximum linear speed of $r\in[0,1]$3 m/s and maximum angular speed of $r\in[0,1]$4 rad/s. Each rollout terminates when the robot reaches the goal, collides with an obstacle, or reaches $r\in[0,1]$5 simulation steps, corresponding to $r\in[0,1]$6 s (Chen et al., 22 Mar 2026).

A practical loop consistent with the method is:

1. Run the low-level controller at $r\in[0,1]$7 Hz.
2. Every $r\in[0,1]$8 control steps, capture the RGB frame and send it asynchronously to the VLM service.
3. Receive a JSON value $r\in[0,1]$9, and clip or rescale if needed.
4. Convert risk to $\alpha_{\text{vlm}}$0 using the bounded power-law map.
5. At every control step, compute nearest-obstacle clearance, normalized clearance, distance cap, speed factor, and $\alpha_{\text{vlm}}$1.
6. Check staleness.
7. Fuse $\alpha_{\text{vlm}}$2 and $\alpha_{\text{vlm}}$3 according to freshness.
8. Compute $\alpha_{\text{vlm}}$4, solve the CBF-QP with $\alpha_{\text{vlm}}$5, and convert $\alpha_{\text{vlm}}$6 to differential-drive commands.

6. Experimental evaluation and comparative behavior

The evaluation covers four scenarios: single frontal obstacle, cluttered field, dynamic obstacle crossing, and dynamic frontal obstacle. These are tested in two environments, a toy flat-ground environment and a realistic warehouse environment. Obstacles include static cylinders of various radii and dynamic obstacles that either oscillate laterally or move directly toward the robot. Each scenario is run $\alpha_{\text{vlm}}$7 times and average results are reported (Chen et al., 22 Mar 2026).

The principal baselines and ablations are: AlphaAdj with no dynamic cap, fixed $\alpha_{\text{vlm}}$8, and fixed $\alpha_{\text{vlm}}$9. The fixed-$\alpha$00 sweep in the warehouse cluttered-field scenario illustrates the tradeoff emphasized by the paper: $\alpha$01 is conservative and increases path length and time to goal, while $\alpha$02 leads to collision (Chen et al., 22 Mar 2026).

Across all $\alpha$03 environment-scenario pairs, AlphaAdj reaches the goal in $\alpha$04 setups with $\alpha$05 collisions. The only other setting that consistently avoids collisions is fixed $\alpha$06, but that conservative baseline sacrifices efficiency and, in one warehouse dynamic obstacle case, fails to reach the goal because it is too cautious. Averaged over all eight setups, fixed $\alpha$07 increases path length by $\alpha$08 meters, reported as $\alpha$09, and increases time-to-goal by $\alpha$10 seconds, or $\alpha$11, relative to AlphaAdj. Fixed $\alpha$12 is too aggressive, producing $\alpha$13 collisions and only $\alpha$14 successful runs overall (Chen et al., 22 Mar 2026).

Compared with the no-cap adaptive variant, AlphaAdj achieves better robustness and slightly better average efficiency. Removing the cap yields $\alpha$15 collisions and only $\alpha$16 successes overall. Averaged across scenarios, AlphaAdj improves over no-cap by $\alpha$17 seconds in time to goal, reported as $\alpha$18. The paper also notes a specific warehouse frontal scenario in which no-cap arrives $\alpha$19 seconds faster and follows a path $\alpha$20 meters shorter than AlphaAdj; however, this is presented as a favorable tradeoff because capping substantially increases robustness against stale and anomalous VLM outputs (Chen et al., 22 Mar 2026).

The warehouse quantitative results include the following examples.

Scenario	AlphaAdj	Comparison
Frontal obstacle	Time-to-goal $\alpha$21 s, minimum margin $\alpha$22 m, path length $\alpha$23 m	No-cap succeeds at $\alpha$24 s and $\alpha$25 m; fixed $\alpha$26 is slower at $\alpha$27 s and $\alpha$28 m; fixed $\alpha$29 fails
Cluttered field	Succeeds with $\alpha$30 s and path length $\alpha$31 m	No-cap fails; fixed $\alpha$32 succeeds but takes $\alpha$33 s and $\alpha$34 m; fixed $\alpha$35 fails
Dynamic obstacle crossing	Succeeds with $\alpha$36 s and path $\alpha$37 m	No-cap succeeds but is slower at $\alpha$38 s and $\alpha$39 m; fixed $\alpha$40 fails to reach destination despite path $\alpha$41 m; fixed $\alpha$42 succeeds with minimum margin $\alpha$43 m
Dynamic frontal obstacle	Succeeds with $\alpha$44 s and path $\alpha$45 m	No-cap fails; fixed $\alpha$46 succeeds at $\alpha$47 s and $\alpha$48 m; fixed $\alpha$49 fails

These results support the paper’s main claim that dynamic semantic modulation of $\alpha$50 gives a better operating point than globally conservative or globally aggressive fixed settings, and that latency-aware capping and fusion are material to reliability.

7. Interpretation, scope, and limitations

AlphaAdj’s conceptual novelty lies in linking semantic scene understanding to a mathematically structured safety mechanism through a bounded scalar risk and an explicit risk-to-$\alpha$51 map. The VLM does not directly issue actions or override safety; it only adjusts the parameter inside a CBF safety filter. This suggests a hybrid architecture in which high-level semantic perception modulates, but does not replace, a certified-style low-level safety mechanism (Chen et al., 22 Mar 2026).

A common misconception would be to treat AlphaAdj as a learned end-to-end navigation controller. The method as described does not do that. Control actions are still produced by a nominal controller plus a CBF-QP safety filter. The semantic model contributes only to the online selection of the conservativeness parameter. A second misconception would be to interpret the VLM risk as a formally calibrated collision probability. The paper explicitly avoids that interpretation and presents the score as a practical proxy for hazard or scene difficulty (Chen et al., 22 Mar 2026).

The method is also careful about latency. Because VLM inference is slower than the control loop, naive use of raw VLM outputs can be dangerous. The uncapped ablation exists precisely to show that latency and occasional outlier VLM estimates can undermine safety if used without the cap and staleness-gated fallback. In this sense, the dynamic cap is not an auxiliary heuristic but an integral part of the proposed system.

The stated limitations are substantial. The evaluation is primarily in simulation and in a constrained set of environments and dynamics. Real-world deployment may introduce sensing noise, actuation delays, communication issues, or more diverse obstacle geometries. The method depends on VLM quality and calibration; if the risk estimate is systematically poor, semantic adaptation may not help, although the cap mitigates worst cases. The risk score is heuristic and prompt-derived rather than a formally grounded probabilistic risk measure. The system also assumes access to geometric obstacle information sufficient to compute barrier values and clearance margins, even though the semantic adaptation uses only RGB. Finally, the paper does not develop a formal closed-loop theorem for delayed, approximate, semantically inferred parameter modulation (Chen et al., 22 Mar 2026).

Within those limits, AlphaAdj defines a specific and technically structured answer to the problem of context-sensitive CBF conservativeness: it uses egocentric RGB observations, asynchronous VLM risk estimation, bounded risk-to-$\alpha$52 modulation, speed- and geometry-aware permissiveness capping, and stale-prediction fallback to make a standard CBF-QP navigation stack more adaptive without discarding its safety-filtering structure (Chen et al., 22 Mar 2026).