cs.ROMar 2, 2026

CHOP: Counterfactual Human Preference Labels Improve Obstacle Avoidance in Visuomotor Navigation Policies

Authors: Gershom Seneviratne, Jianyu An, Vaibhav Shende, Sahire Ellahy, Yaxita Amin, Kondapi Manasanjani, Samarth Chopra, Jonathan Deepak Kannan, 

Organizations: University of Maryland, College Park

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Visuomotor navigation policies have shown strong perception-action coupling for embodied agents, yet they often struggle with safe navigation and dynamic obstacle avoidance in complex real-world environments. We introduce CHOP, a novel approach that leverages Counterfactual Human Preference Labels to align visuomotor navigation policies towards human intuition of safety and obstacle avoidance in navigation. In CHOP, for each visual observation, the robot's executed trajectory is included among a set of counterfactual navigation trajectories: alternative trajectories the robot could have followed under identical conditions. Human annotators provide pairwise preference labels over these trajectories based on anticipated outcomes such as collision risk and path efficiency. These aggregated preferences are then used to fine-tune visuomotor navigation policies, aligning their behavior with human preferences in navigation. Experiments on the SCAND dataset show that visuomotor navigation policies fine-tuned with CHOP reduce near-collision events by 49.7%, decrease deviation from human-preferred trajectories by 45.0%, and increase average obstacle clearance by 19.8% on average across multiple state-of-the-art models, compared to their pretrained baselines. These improvements transfer to real-world deployments on a Ghost Robotics Vision60 quadruped, where CHOP-aligned policies improve average goal success rates by 24.4%, increase minimum obstacle clearance by 6.8%, reduce collision and intervention events by 45.7%, and improve normalized path completion by 38.6% on average across navigation scenarios, compared to their pretrained baselines. Our results highlight the value of counterfactual preference supervision in bridging the gap between large-scale visuomotor policies and human-aligned, safety-aware embodied navigation.

Figures & tables

Fig. 2: Overview of CHOP. (1) Counterfactual dataset generation: Given a navigation dataset containing egocentric images and the trajectories executed by a robot, we generate multiple counterfactual trajectory alternatives for each observation. Human annotators provide binary preference labels over pairs of counterfactual trajectories, resulting in multiple comparisons per image (horizontal axis), while the process is repeated across the entire dataset (vertical axis). The aggregated preference annotations form the CHOP counterfactual preference dataset. (2) Policy training: For each observation, the dataset with the most preferred trajectory based on multiple preference labels for the same observation ( * ), is extracted from the binary rankings and used to fine-tune pretrained visuomotor navigation policies (e.g., OmniVLA [ 12 ] , ViNT [ 29 ] , GNM [ 28 ] ) using Supervised Fine-Tuning (SFT) or Low Rank Adaptation (LoRA), producing a preference-aligned visuomotor policy. (3) Deployment: The resulting preference-aligned policies are deployed on real robot platforms, where predicted paths are executed by a downstream controller. Evaluations on both offline datasets and real-world robots demonstrate the benefits of counterfactual preference supervision for safer and more human aligned navigation.
Fig. 3: Counterfactual trajectory generation and human preference annotation under identical egocentric observations. A : Trajectory recorded in the dataset. B–C : Structured counterfactual trajectories generated by rotating the executed trajectory counterclockwise and clockwise respectively. D : Human-guided target trajectory. E : Preference annotation interface, where annotators compare two overlaid trajectories and select the safer or more appropriate one. Trajectories are rendered in the robot’s egocentric frame; colors indicate different candidates and do not imply preference.
Number of annotated observations187,920
Counterfactuals per observation ( MM )4
Total counterfactual trajectories751,680
Fraction where dataset trajectory is not preferred70.51%
Total pairwise comparisons1,127,520
TABLE I: Statistics of the CHOP preference dataset.
Fig. 4: Qualitative comparison of baseline and CHOP-finetuned visuomotor policies across three navigation scenarios, with rows corresponding to methods and columns corresponding to scenarios. Green trajectories denote CHOP-finetuned policies, while red dashed trajectories denote baseline behavior. GNM [ 28 ] and OmniVLA [ 12 ] exhibit improved obstacle avoidance and increased clearance after fine-tuning, whereas ViNT [ 29 ] shows more limited qualitative change.
ModelNear-Coll. ↓\downarrowDeviation ↓\downarrowClearance ↑\uparrow
OmniVLA (Base)24530.63452.7647
OmniVLA + CHOP897 (-63.4%)0.3185 (-49.5%)3.0332 (+9.7%)
ViNT (Base)51921.80822.2472
ViNT + CHOP3848 (-25.9%)1.3216 (-26.9%)2.5065 (+11.5%)
GNM (Base)55291.94742.1629
GNM + CHOP2219 (-59.7%)0.8056 (-58.6%)2.9887 (+38.1%)
TABLE II: Offline evaluation on SCAND. CHOP-aligned models demonstrate improved average obstacle clearance and reduced near-collision counts while reducing average deviation from the expert annotation for preferred trajectories.
MetricsModelScenario 1Scenario 2Scenario 3
Success Rate (%) ↑\uparrowOmniVLA (Base)0.080.00.0
OmniVLA + CHOP20.0100.0100.0
ViNT (Base)0.00.040.0
ViNT + CHOP0.00.020.0
GNM (Base)60.00.00.0
GNM + CHOP100.00.060.0
TABLE III: Real-world evaluation across three navigation scenarios. CHOP-aligned models consistently improve safety-related metrics while preserving task success across diverse environments.
ParameterValue
Total steps12,500
Batch size per GPU4 (grad accum. 10)
Learning rate1×10−41\times 10^{-4}
LR decay start100,000 steps (10 ×\times drop)
Waypoints per sample8
LoRA rank32
TABLE IV: OmniVLA fine-tuning hyperparameters.
SettingValue
Epochs30
Batch size per GPU1024
Learning rate5×10−45\times 10^{-4}
OptimizerAdamW
SchedulerCosine (warmup 4 epochs)
Context size5 frames (temporal)
TABLE V: ViNT fine-tuning hyperparameters.
ParameterValue
Epochs40
Batch size per GPU2048
Learning rate7×10−47\times 10^{-4}
OptimizerAdam
Context size5 frames (temporal)
Predicted horizon5 waypoints
TABLE VI: GNM fine-tuning hyperparameters.
Fig. 5: Offline qualitative comparison on the SCAND test set showing successive frames (left to right) from three navigation scenarios. Each scenario is visualized across three rows corresponding to OmniVLA, GNM, and ViNT. Blue trajectories denote baseline model predictions, while red trajectories denote CHOP-aligned predictions. When a trajectory disappears from view, the policy has slowed or stopped due to conservative planning. Across most scenarios, timestamps, and methods, CHOP-aligned policies exhibit earlier obstacle avoidance, increased clearance, and more cautious behavior in the presence of dynamic agents compared to their baseline counterparts.
Fig. 6: Distribution of offline evaluation metrics on the SCAND test set for baseline (yellow) and CHOP-finetuned (pink) policies across GNM, ViNT, and OmniVLA. From top to bottom: near-collision count, deviation from human-preferred trajectories, and minimum obstacle clearance. CHOP fine-tuning shifts the distributions toward fewer near-collisions, lower deviation, and increased clearance, indicating improved safety and alignment beyond mean performance.
Fig. 7: ROS 2-based deployment architecture used for real-world execution. A model runner asynchronously produces short-horizon waypoint sequences, a path manager maintains and updates the active trajectory, and a planner tracks the current waypoint to generate velocity commands. The design decouples policy inference from low-level control and supports modular swapping of visuomotor policies.
More in this topic

Safe Robot Navigation

Safe robot navigation develops methods for reaching goals while avoiding collisions and managing uncertainty, dynamic obstacles, and environmental hazards.

See all 139 papers

Explore similar work

CardsList
Jun 10, 2026

SAFER-Nav: Enhancing Safety for Visual Robot Navigation via Segmentation-Aware Fine-Tuning

Vision-based navigation models, particularly foundation models, generate viable trajectories from RGB observations alone. However, even state-of-the-art transformer- and diffusion-based policies struggle to generalize in unfamiliar deployment environments containing unseen obstacles or shifted conditions. The resulting trajectories often remain goal-directed but unsafe. Existing efforts improve safety through external trajectory correction or internal geometric priors, yet the resulting policies are not trained to explicitly represent obstacle boundaries or traversable free-space structure. To address this, we propose a navigation model that incorporates these structures directly into the policy via fine-tuning and is designed to be compatible with diverse RGB-based backbones. Across multiple robot platforms, indoor environments, and static and dynamic obstacle scenarios, our method reduces collision frequency relative to ViNT, NoMaD, and their CARE-augmented variants while maintaining goal-reaching performance.

Geonyeong Ko, Giung Lee, Changjoo Nam

Dept. of Electronic Engineering, Sogang University, Seoul, Korea · Dept. of Computer Science, Rice University, Houston, TX, USA · Vertical Labs, Co., Ltd., Seoul, Korea

Robot Policy LearningTrajectory GenerationRobot NavigationSafe Robot NavigationRobot SafetyVisual Robot Navigation
AskPDF
Jul 16, 2026

NavCMPO: Critic-Guided MeanFlow Policy Optimization for Adaptive Navigation

End-to-end diffusion-based policies have demonstrated strong performance in mapless visual navigation, but their iterative denoising process introduces substantial inference latency, while behavior cloning limits performance to the quality of expert demonstrations. We present NavCMPO, a two-stage adaptive navigation framework that combines few-step MeanFlow trajectory generation, critic-guided refinement, and reinforcement learning fine-tuning. During pre-training, an obstacle proximity prediction task encourages the visual representation to capture obstacle-aware spatial information. To compensate for the degradation in obstacle avoidance caused by few-step generation, Critic-Guided Trajectory Refinement (CGTR) uses gradients from a critic trained with obstacle-point-cloud supervision to refine intermediate trajectories. During adaptation, the MeanFlow policy is fine-tuned using Proximal Policy Optimization with behavior-cloning regularization, while the critic is updated to accommodate embodiment-specific observation changes. Under a matched training budget on the InternVLA-N1 benchmark, NavCMPO achieves an average success rate of 74.7%, exceeding the retrained NavDP baseline by 6.4 percentage points, while reducing inference latency from 85,ms to 60,ms. Experiments on a Unitree Go2 further demonstrate effective sim-to-real transfer.

Junjie An, Yi Wu, Xiao Liu +5

Zhejiang University, Hangzhou, China · Shandong University

Flow MatchingRobot Policy AdaptationTrajectory GenerationMeanFlowRobot NavigationDiffusion Policy+1
AskPDF
Oct 26, 2026

Benchmarking Obstacle-Conditioned Safety in Household Robot Navigation

Safe household navigation requires robots to regulate clearance and speed according to what an obstacle is and how close it is, without unnecessary conservatism. Yet existing evaluations do not test whether semantic safety distinctions are reflected in continuous motion. To that end, we introduce OSMA-Bench, an Obstacle-Conditioned Safe Motion Adaptation Benchmark built on spatially matched scenario sets that differ only in obstacle type. We also propose the Semantic Sensitivity Index (SSI), which measures whether clearance and distance-dependent speed and smoothness follow the caution ordering of obstacles. Evaluating VLM-based and vision-language-action (VLA) policies on the benchmark, we find that neither consistently adapts its motion to obstacle safety, even with explicit safety prompts. Our analysis shows that the policies can correctly distinguish obstacles by safety relevance but fail to reflect this distinction in the motion parameters or actions they produce. For VLM-based policies, which recognize the obstacle and then generate a navigation program, adding an explicit parameter-inference step between these two stages improves adaptation. These findings expose a gap in semantic-to-motion grounding, highlighting the need to translate semantic caution into appropriately safe motion.

Hyunwoong Bae, Jisu Han, Jiyeon Choi +2

Interdisciplinary Program in Artificial Intelligence, Seoul National University · Department of Electrical and Computer Engineering, Seoul National University

Safe Robot NavigationRobot Policy AdaptationRobot Navigation BenchmarksSafety Evaluation of VLA Models
AskPDF