Less Is More: Scalable Visual Navigation from Limited Data

TL;DR

LiMo addresses data efficiency in goal-conditioned visual navigation by augmenting limited expert demonstrations with planner-generated trajectories. It introduces a transformer-based policy that, from a single RGB frame and an SE(2) goal, predicts a sequence of $N$ waypoints over a horizon $T$, trained on a mix of teleoperation data and MPPI-generated supervision. Experiments on GrandTour show that geometric augmentation yields substantial improvements in SPL for diverse, obstacle-rich environments and that LiMo learns embodiment-specific navigation for the ANYmal D quadruped. The work demonstrates practical real-robot deployment and argues that strategically curated, high-quality geometric supervision can be more effective than merely collecting more demonstrations for scalable, data-efficient visual navigation.

Abstract

Imitation learning provides a powerful framework for goal-conditioned visual navigation in mobile robots, enabling obstacle avoidance while respecting human preferences and social norms. However, its effectiveness depends critically on the quality and diversity of training data. In this work, we show how classical geometric planners can be leveraged to generate synthetic trajectories that complement costly human demonstrations. We train Less is More (LiMo), a transformer-based visual navigation policy that predicts goal-conditioned SE(2) trajectories from a single RGB observation, and find that augmenting limited expert demonstrations with planner-generated supervision yields substantial performance gains. Through ablations and complementary qualitative and quantitative analyses, we characterize how dataset scale and diversity affect planning performance. We demonstrate real-robot deployment and argue that robust visual navigation is enabled not by simply collecting more demonstrations, but by strategically curating diverse, high-quality datasets. Our results suggest that scalable, embodiment-specific geometric supervision is a practical path toward data-efficient visual navigation.

Figures (9)

• Figure 1: Starting from GrandTour GrandTour, a high-quality dataset of recorded quadruped robot missions, we augment the limited real-world expert demonstrations with planner-generated trajectories to scale supervision for imitation learning. The augmented dataset is then used to train LiMo, an end-to-end visual navigation policy.
• Figure 2: Three representative samples from the augmented dataset $\mathcal{D}_\text{AUG}$. Each row shows the front camera image (left) and the corresponding elevation map (right). Teleoperated, geometric, and real-world paths are overlaid with their associated goals.
• Figure 3: Overview of LiMo's architecture. The policy takes as input a single RGB image $\mathbf{I}$, and a goal pose $\mathbf{g} = (x, y, \theta)$ in the robot-centric frame. Image features are extracted using a frozen DINOv2 encoder and combined with learned positional embeddings. A transformer decoder conditioned on the goal embedding predicts $N$ waypoint embeddings, which are linearly projected to $N$ waypoints $(x, y, \theta)$, forming the output trajectory.
• Figure 4: Qualitative comparison of predicted trajectories across diverse indoor and outdoor scenarios. Images (b, c, e, f, h, i) are included in the training set, while (a, d, g, j, k, m) are held-out scenes. Column three compares predictions from the teleop-only policy with those from the augmented-trained policy, where the latter shows much better scene understanding. Column four illustrates the remaining failure cases. LiMo correctly recognizes door openings (a, j), handles stair ascent and descent consistent with ANYmal’s embodiment (e, i), and navigates unstructured natural terrain such as forests and fields (c, f). It also demonstrates robustness on construction sites (g) and in high-elevation outdoor environments (k), while remaining challenged by cliff edges (d), transparent surfaces, and obstacles such as glass doors and nets (h, m), and by unusual obstacles such as snow piles (m).
• Figure 5: (Left) Evaluation of SPL on $\mathcal{D}_\text{TEL}$ and $\mathcal{D}_\text{AUG}$ test sets trained on a varying number of augmenting geometric paths. Adding geometric samples mainly benefits performance on the more diverse $\mathcal{D}_\text{AUG}$ test split. (Middle) When evaluating goals following the $\mathcal{D}_\text{AUG}$ distribution, models trained solely on $\mathcal{D}_\text{TEL}$ consistently underperform compared to those trained on the $\mathcal{D}_\text{AUG}$. While training on $\mathcal{D}_\text{AUG}$ generalizes well to $\mathcal{D}_\text{TEL}$-style goals, models trained only on $\mathcal{D}_\text{TEL}$ fail to generalize to unseen diverse goals with limited data. (Right) Evaluation of SPL on $\mathcal{D}_\text{TEL}$ goals trained using the $\mathcal{D}_\text{TEL}$ and $\mathcal{D}_\text{AUG}$. Training on just 5 missions of $\mathcal{D}_\text{AUG}$ achieves comparable performance to training on $\mathcal{D}_\text{TEL}$ with 41 missions.