RVN-Bench: A Benchmark for Reactive Visual Navigation

AI summary

Overview

Research area: Robotics — specifically indoor mobile robot navigation, embodied AI simulation benchmarks, reinforcement learning, safe reinforcement learning, and imitation learning.

Technical level: Intermediate. The paper is readable without deep RL background, but understanding the baseline comparisons benefits from familiarity with standard navigation metrics (success rate), policy-gradient RL (PPO, DD-PPO), and imitation-learning navigation models (ViNT, NoMaD).

Scope (one sentence): The paper introduces RVN-Bench, a simulated benchmark built on Habitat 2.0 with HM3D scenes that measures whether indoor mobile robots can reach a sequence of goal positions from first-person RGB images alone while avoiding all collisions.

What This Paper Is About

Reactive visual navigation means reaching a goal using only what the camera currently sees, with no pre-built map and no task-specific knowledge. Most existing navigation benchmarks either score only whether the robot reached the goal and ignore whether it crashed into walls or furniture, or they are built for outdoor driving rather than cluttered indoor spaces. The authors build a benchmark where collisions are explicitly detected, penalized, and reported, and where the robot must survive a chain of successive goals rather than a single destination.

Key Contributions

1. RVN-Bench itself: a collision-aware evaluation framework for reactive visual navigation, built on Habitat 2.0 using HM3D scenes, with a defined task (reach sequential goals from visual observations only) and collision-aware metrics. It uses 800 scenes for training, 50 for validation, and 50 for testing.
2. An RL training environment: an interactive environment for online reinforcement learning, with goal-reaching rewards plus an explicit collision penalty, and an additional collision cost for Safe-RL training.
3. An offline trajectory image dataset generator: a pipeline producing expert (collision-free) trajectory image datasets in GNM dataset format, plus a negative trajectory image dataset generator that captures trajectories ending in collisions — data that is expensive and unsafe to collect on real robots.
4. A baseline suite and a new baseline (NoMaD-Neg): evaluations across RL (PPO, DD-PPO), Safe-RL (PPO-Lagrangian), and IL methods (ViNT-PointGoal, NoMaD-PointGoal), plus NoMaD-Neg, which uses the benchmark's negative trajectory datasets to improve collision avoidance.

Main Findings

• Collision-aware visual navigation is far from solved. The best baseline, DDPPO-DAV2, reached a test success rate at the first goal (SR₁) of 0.928 and 20.79 average goals per episode E(G) with 3.6 collisions per kilometer (CPK). The paper states these results indicate the problem remains unsolved, and notes the difficulty is much greater than for visual navigation tasks that ignore collisions.
• Depth estimates from a foundation model help substantially. DDPPO-DAV2, which augments RGB with depth predicted by Depth Anything V2 (DAV2), achieved 1.67× and 1.49× higher E(G) than DD-PPO on the validation and test sets, and reduced CPK by 62% and 59%.
• RL beat imitation learning across all metrics. All RL-based methods outperformed IL methods. The ResNet-LSTM used by the RL methods has fewer parameters (12M) than ViNT-PointGoal (23M) and NoMaD-PointGoal (13M), yet performed better, suggesting that learning through interaction with the environment matters more than model size here.
• ViNT-PointGoal struggled badly. Its test SR₁ was 0.093, E(G) 0.10, and CPK 465.4, while NoMaD-PointGoal reached test SR₁ 0.751, E(G) 4.52, CPK 31.0.
• Negative data improved the imitation baseline in unseen scenes. NoMaD-Neg outperformed NoMaD-PointGoal on all metrics in validation and test — test SR₁ 0.760 vs. 0.751, E(G) 4.61 vs. 4.52, CPK 25.8 vs. 31.0 — but still underperformed every RL method on all metrics.
• Generalization to unseen scenes was strong. SR₁ on validation and test scenes was on average about 1.5% lower than on training scenes.
• Ground-truth depth beats predicted depth. In the ablation, DD-PPO with RGB plus ground-truth depth achieved an average 10.2% improvement in E(G) over predicted depth on validation and 10.6% on test, with CPK decreasing by roughly 45% (validation) and 50% (test). The authors read this as evidence that better monocular depth estimation would further improve reactive navigation.
• Promising sim-to-real transfer in a small physical study. A NoMaD-PointGoal model trained only on RVN-Bench simulation data achieved a real-world E(G) 3.5× higher than a model trained only on real data (1.05 vs. 0.30), while a model trained on both real and simulation data achieved a 4.3× improvement (1.30). CPK was 223.5 for the real-only model, 196.4 for the simulation-only model, and 191.4 for the combined model.

Methodology in Plain English

The task. An agent starts at a random pose in a scene it has never seen and receives RGB images from a forward-facing camera plus the goal's position relative to itself. It picks from four discrete actions: move forward by 0.25 m, turn left 30 degrees, turn right 30 degrees, or stop. If it stops within twice its radius of the goal, a new goal is issued without resetting the scene. An episode ends in one of three ways: the agent collides, it runs out of the maximum step budget (T_max), or it reaches all N_goal sequential goals. The robot is a cylinder of radius 0.18 m and height 1.0 m with the camera at 0.6 m; images are 256×256 and goals are sampled at geodesic distances between 4 m and 8 m.

Collision handling. Collisions are checked against a precomputed navigation mesh (NavMesh) built from scene geometry. On a forward move, the target position is projected onto the NavMesh while respecting the robot's radius. If the resulting displacement is smaller than the commanded step, the move was blocked and a collision is registered. The paper describes this as a hard kinematic constraint — no physics-based contact dynamics are simulated.

Rewards. Reaching a goal gives +1.0; colliding gives a terminal −0.1; all other steps give the change in geodesic distance to the goal minus 0.01. Safe-RL adds a cost of 1.0 on collision.

Data generation. For expert data, start and goal are sampled, an A* shortest path is computed on the ground-truth occupancy map padded by more than the agent radius (so the path is safe for discrete actions), and the agent's positions, yaws, and images are recorded if it reaches the goal without collision. For negative data, the pad margin is set smaller than the agent radius so the path is unsafe; when the agent collides at t_c, a window from t_i = max(0, t_c − k_pre) to t_f = min(T, t_c + k_post) is recorded, with k_pre = 8 and k_post = 6, and the image frozen at the collision frame for all later steps.

Training setup. IL baselines were trained for 100 epochs on data from 800 scenes at 400 trajectories per scene (320k trajectories total, roughly 828 hours at 1.0 m/s). NoMaD-Neg used the same trajectory budget but replaced 6.25% of expert trajectories with negative ones, and at inference samples 8 trajectories from each of its expert and negative models, picking the expert trajectory that minimizes a constrained reward (CoR) relative to the negative set. RL and Safe-RL were trained until convergence; PPO-Lagrangian used an expected cost limit of 10⁻⁴. Everything ran on a single NVIDIA L40S GPU except DD-PPO and DDPPO-DAV2, which used 8 L40S GPUs.

Metrics. Because new goals keep arriving, the paper reports SR₁ (success rate on the first goal), E(G) (average number of goals reached per episode), and CPK (collisions per kilometer traveled).

Real-world test. A Clearpath Jackal UGV with a ZED-X 2.2mm camera at 0.6 m and an onboard NVIDIA Jetson AGX Orin ran 20 episodes (15 in offices, 5 in houses), each with N_g = 5 goals, using the same discrete action set, with a goal counted as reached within 0.36 m.

Why This Matters

Impact on research. Earlier indoor benchmarks such as the Habitat Challenge and GOAT-Bench ignore collisions entirely, and HabiCrowd only accounts for pedestrians, not static obstacles like walls and furniture. Collision-aware benchmarks that do exist — Safety-Gym and BARN — rely on LiDAR rather than vision, and autonomous-driving benchmarks like CARLA and SUMMIT are outdoor-specific. RVN-Bench occupies an otherwise empty slot: indoor, vision-only, collision-aware, with a lifelong sequence of goals. It also provides a safe way to generate collision data, which cannot easily be collected on real hardware.

Real-world applications:

• Indoor service and delivery robots navigating offices, hotels, and hospitals where furniture and walls must be avoided.
• Warehouse and logistics robots operating around static shelving and equipment.
• Home assistant robots moving between rooms in cluttered domestic environments.
• Inspection or telepresence robots that must operate safely around people and property without a pre-built map.

Industry relevance. The paper's sim-to-real experiment suggests that large-scale simulation datasets (828 hours generated from RVN-Bench) can complement limited and expensive real-world data — the combined dataset gave the best real-world results in the study. That has direct cost implications for companies that cannot afford extensive real-world data collection, and the benchmark's configurable robot footprint, motion discretization, and camera parameters make it adaptable to different hardware. The gap between simulation success (DDPPO-DAV2 at 0.928 test SR₁) and real-world success (best real-world SR₁ of 0.75 with only 5 goals per episode) also gives industry a concrete sense of how much safety headroom remains.

Future Directions

• Dynamic obstacles. The benchmark is limited to static environments; extending it to moving obstacles such as people is stated as future work.
• Continuous control. The current action space is discrete (forward step and 30-degree turns); the authors plan continuous control instead.
• Broader robot support and larger physical evaluation. The real-world validation used only a single hardware platform (the Jackal UGV) and, by the authors' own description, was limited in scale.
• Closing the depth gap. Because ground-truth depth clearly beat predicted depth on both E(G) and CPK, the authors expect further advances in monocular depth estimation to translate into better reactive visual navigation — and the benchmark provides a way to measure that.

Target Audience

Robotics and embodied-AI researchers who need a standardized, collision-aware testbed for indoor visual navigation; reinforcement-learning and safe-RL researchers working on navigation policies; imitation-learning researchers working on goal-conditioned navigation foundation models such as ViNT and NoMaD; and applied engineers at robotics companies who need to compare navigation policies on safety as well as goal-reaching, or who want to supplement scarce real-world data with simulation trajectories.

Authors’ abstract

Safe visual navigation is critical for indoor mobile robots operating in cluttered environments. Existing benchmarks, however, often neglect collisions or are designed for outdoor scenarios, making them unsuitable for indoor visual navigation. To address this limitation, we introduce the reactive visual navigation benchmark (RVN-Bench), a collision-aware benchmark for indoor mobile robots. In RVN-Bench, an agent must reach sequential goal positions in previously unseen environments using only visual observations and no prior map, while avoiding collisions. Built on the Habitat 2.0 simulator and leveraging high-fidelity HM3D scenes, RVN-Bench provides large-scale, diverse indoor environments, defines a collision-aware navigation task and evaluation metrics, and offers tools for standardized training and benchmarking. RVN-Bench supports both online and offline learning by offering an environment for online reinforcement learning, a trajectory image dataset generator, and tools for producing negative trajectory image datasets that capture collision events. Experiments show that policies trained on RVN-Bench generalize effectively to unseen environments, demonstrating its value as a standardized benchmark for safe and robust visual navigation. Code and additional materials are available at: https://rvn-bench.github.io/.