MINav: Efficient Image-Goal Navigation

• MINav is an image-goal navigation method that formulates the problem as an offline goal-conditioned reinforcement learning task using unsupervised data collection and hindsight goal relabeling.
• It leverages frozen DINOv3 visual features with a sparse similarity-based reward and an offline TD3+BC learning procedure to efficiently learn compact navigation policies.
• The method enables rapid real-world deployment in under 120 minutes on consumer hardware while outperforming several zero-shot navigation baselines.

MINav is an image-goal visual navigation method that formulates ImageNav as an offline goal-conditioned reinforcement learning problem, combining unsupervised data collection with hindsight goal relabeling and offline policy learning. Its central claim is that a dataset can be collected, an in-domain policy trained, and the result deployed in the real world in less than 120 minutes, on a consumer laptop, and without human intervention (Liu et al., 27 Mar 2026). Within that formulation, MINav uses frozen visual features, sparse goal achievement defined by representation similarity, and an offline TD3+BC-style learning procedure to produce a compact policy that, in the reported experiments, outperforms several zero-shot navigation baselines in target environments and scales favorably with dataset size (Liu et al., 27 Mar 2026).

1. Problem formulation

MINav models image-goal navigation as a goal-conditioned partially observable Markov decision process. The state is a history of visual embeddings,

$$s_t \rightsquigarrow [o'_{t-3}, o'_{t-2}, o'_{t-1}, o'_t] \in \mathbb{R}^{4 \times d},$$

where $o'_t=\phi(o_t)\in\mathbb{R}^d$ is the frozen DINOv3 feature of the raw RGB image $o_t$. The observation space is RGB imagery, the action space is

$\mathcal{A}=[v_x,v_y,\omega_z]\in[-1,1]^3,$

and transitions are induced by robot dynamics together with collision avoidance via LiDAR (Liu et al., 27 Mar 2026).

The goal space is constructed from valid goal embeddings filtered by a spatial-standard-deviation threshold,

$$\mathcal{G}=\{\phi(o)\mid \sigma_{\mathrm{spa}}(o)>\delta_{\mathrm{ssd}}\}, \qquad \delta_{\mathrm{ssd}}=0.02.$$

Goal achievement is defined through a sparse reward based on state-goal similarity:

$$R(s_t,g)=1 \;\; \text{if} \;\; S(s_t,g)\ge \delta_{\mathrm{done}} \;\; \text{else} \;\; 0,$$

with $\delta_{\mathrm{done}}=0.8$ and

$S(s_t,g)=\frac{1}{4}\sum_{k=t-3}^{t}\cos(o'_k,g).$

The discount factor is $\gamma=0.99$, and the horizon is fixed as $H\to\infty$ until “done” (Liu et al., 27 Mar 2026).

This formulation makes the representation layer structurally central. A plausible implication is that navigation success depends not only on control learning, but also on the geometry induced by the frozen visual encoder and on the filtering of admissible goals through $o'_t=\phi(o_t)\in\mathbb{R}^d$0.

2. Unsupervised exploration and dataset construction

MINav begins from an unlabeled offline dataset,

$o'_t=\phi(o_t)\in\mathbb{R}^d$1

with no human labels or goals. The data-collection policy is “Pink Uniform Noise,” introduced to maximize state-action coverage under at most two hours of real-robot runtime (Liu et al., 27 Mar 2026).

The procedure generates temporally correlated pink noise in the frequency domain by creating a white-noise spectrum, shaping its magnitude proportionally to $o'_t=\phi(o_t)\in\mathbb{R}^d$2 with $o'_t=\phi(o_t)\in\mathbb{R}^d$3, and applying an inverse FFT to obtain a Gaussian-distributed sequence $o'_t=\phi(o_t)\in\mathbb{R}^d$4. The marginals are then converted to uniform variables through

$o'_t=\phi(o_t)\in\mathbb{R}^d$5

where $o'_t=\phi(o_t)\in\mathbb{R}^d$6 is the standard-normal cumulative distribution function, and finally linearly rescaled to the control range for each action dimension (Liu et al., 27 Mar 2026).

Execution occurs at $o'_t=\phi(o_t)\in\mathbb{R}^d$7 and is upsampled to $o'_t=\phi(o_t)\in\mathbb{R}^d$8 through smoothing. RGB frames are recorded at $o'_t=\phi(o_t)\in\mathbb{R}^d$9 and resized to $o_t$0 for DINOv3 feature extraction. LiDAR is used only for safety, stopping the robot if an obstacle is too close. The reported data budgets are approximately $o_t$1 steps for one hour and approximately $o_t$2 steps for two hours (Liu et al., 27 Mar 2026).

For supervised targets, MINav relies on hindsight goal relabeling. For each transition $o_t$3, a hindsight goal is sampled according to

$o_t$4

for critic training, while actor updates use $o_t$5. The geometric component samples $o_t$6 with

$o_t$7

then sets $o_t$8. The reward under relabeling retains the same similarity-threshold form (Liu et al., 27 Mar 2026).

The empirical motivation for this design is explicit in the reported simulation results: pink uniform exploration obtains the highest normalized state, action, and joint state-action entropy among the compared noise strategies, and yields better downstream policy success rates than Ornstein-Uhlenbeck, white, and pink-Gaussian alternatives under the stated budgets (Liu et al., 27 Mar 2026).

3. Offline policy learning and model architecture

MINav uses an offline TD3+BC procedure. Two critic networks $o_t$9 minimize

$\mathcal{A}=[v_x,v_y,\omega_z]\in[-1,1]^3,$0

with target

$\mathcal{A}=[v_x,v_y,\omega_z]\in[-1,1]^3,$1

The actor $\mathcal{A}=[v_x,v_y,\omega_z]\in[-1,1]^3,$2 maximizes

$\mathcal{A}=[v_x,v_y,\omega_z]\in[-1,1]^3,$3

with $\mathcal{A}=[v_x,v_y,\omega_z]\in[-1,1]^3,$4. Target networks are updated by Polyak averaging with $\mathcal{A}=[v_x,v_y,\omega_z]\in[-1,1]^3,$5, and exploration noise satisfies $\mathcal{A}=[v_x,v_y,\omega_z]\in[-1,1]^3,$6 (Liu et al., 27 Mar 2026).

Offline model selection is performed with fitted Q-evaluation. A separate critic $\mathcal{A}=[v_x,v_y,\omega_z]\in[-1,1]^3,$7 is trained by minimizing

$\mathcal{A}=[v_x,v_y,\omega_z]\in[-1,1]^3,$8

where

$\mathcal{A}=[v_x,v_y,\omega_z]\in[-1,1]^3,$9

Checkpoints are ranked by the average $$\mathcal{G}=\{\phi(o)\mid \sigma_{\mathrm{spa}}(o)>\delta_{\mathrm{ssd}}\}, \qquad \delta_{\mathrm{ssd}}=0.02.$$0 at initial states (Liu et al., 27 Mar 2026).

The encoder is a frozen DINOv3-ViT-S with $$\mathcal{G}=\{\phi(o)\mid \sigma_{\mathrm{spa}}(o)>\delta_{\mathrm{ssd}}\}, \qquad \delta_{\mathrm{ssd}}=0.02.$$1 parameters, or alternatively ViT-L. The actor is a 3-layer MLP with dimensions $$\mathcal{G}=\{\phi(o)\mid \sigma_{\mathrm{spa}}(o)>\delta_{\mathrm{ssd}}\}, \qquad \delta_{\mathrm{ssd}}=0.02.$$2, ReLU activations, and tanh output. Each critic is a 4-layer MLP with input consisting of state, goal, and action, hidden widths $$\mathcal{G}=\{\phi(o)\mid \sigma_{\mathrm{spa}}(o)>\delta_{\mathrm{ssd}}\}, \qquad \delta_{\mathrm{ssd}}=0.02.$$3, and ReLU nonlinearities. The total policy size is approximately $$\mathcal{G}=\{\phi(o)\mid \sigma_{\mathrm{spa}}(o)>\delta_{\mathrm{ssd}}\}, \qquad \delta_{\mathrm{ssd}}=0.02.$$4 parameters. Optimization uses Adam with learning rate $$\mathcal{G}=\{\phi(o)\mid \sigma_{\mathrm{spa}}(o)>\delta_{\mathrm{ssd}}\}, \qquad \delta_{\mathrm{ssd}}=0.02.$$5, no weight decay, batch size $$\mathcal{G}=\{\phi(o)\mid \sigma_{\mathrm{spa}}(o)>\delta_{\mathrm{ssd}}\}, \qquad \delta_{\mathrm{ssd}}=0.02.$$6, and $$\mathcal{G}=\{\phi(o)\mid \sigma_{\mathrm{spa}}(o)>\delta_{\mathrm{ssd}}\}, \qquad \delta_{\mathrm{ssd}}=0.02.$$7 training steps (Liu et al., 27 Mar 2026).

This architecture is intentionally asymmetric: representation learning is frozen and large, while the control policy is small. This suggests that MINav treats navigation primarily as a problem of exploiting a pretrained perceptual manifold with lightweight task-specific policy fitting, rather than jointly learning perception and control end to end.

4. Computational profile and deployment pipeline

A defining property of MINav is its end-to-end computational economy. The reported hardware is a consumer-grade laptop with an NVIDIA RTX 5070 GPU and $$\mathcal{G}=\{\phi(o)\mid \sigma_{\mathrm{spa}}(o)>\delta_{\mathrm{ssd}}\}, \qquad \delta_{\mathrm{ssd}}=0.02.$$8 RAM. Data collection takes $$\mathcal{G}=\{\phi(o)\mid \sigma_{\mathrm{spa}}(o)>\delta_{\mathrm{ssd}}\}, \qquad \delta_{\mathrm{ssd}}=0.02.$$9 to $$R(s_t,g)=1 \;\; \text{if} \;\; S(s_t,g)\ge \delta_{\mathrm{done}} \;\; \text{else} \;\; 0,$$0 minutes on the robot. Data processing, including feature extraction and SSD filtering, requires $$R(s_t,g)=1 \;\; \text{if} \;\; S(s_t,g)\ge \delta_{\mathrm{done}} \;\; \text{else} \;\; 0,$$1 minutes with ViT-S or $$R(s_t,g)=1 \;\; \text{if} \;\; S(s_t,g)\ge \delta_{\mathrm{done}} \;\; \text{else} \;\; 0,$$2 minutes with ViT-L, while policy training requires $$R(s_t,g)=1 \;\; \text{if} \;\; S(s_t,g)\ge \delta_{\mathrm{done}} \;\; \text{else} \;\; 0,$$3 minutes (Liu et al., 27 Mar 2026).

The timing summary is compact:

Pipeline configuration	Total time
1 h collection + ViT-S	85 min
1 h collection + ViT-L	111 min
2 h collection + ViT-S	148 min
2 h collection + ViT-L	193 min

Only the $$R(s_t,g)=1 \;\; \text{if} \;\; S(s_t,g)\ge \delta_{\mathrm{done}} \;\; \text{else} \;\; 0,$$4-hour collection settings satisfy the “under 120 minutes” criterion, namely $$R(s_t,g)=1 \;\; \text{if} \;\; S(s_t,g)\ge \delta_{\mathrm{done}} \;\; \text{else} \;\; 0,$$5 minutes for ViT-S and $$R(s_t,g)=1 \;\; \text{if} \;\; S(s_t,g)\ge \delta_{\mathrm{done}} \;\; \text{else} \;\; 0,$$6 minutes for ViT-L (Liu et al., 27 Mar 2026). The training cost itself is nearly constant across the reported $$R(s_t,g)=1 \;\; \text{if} \;\; S(s_t,g)\ge \delta_{\mathrm{done}} \;\; \text{else} \;\; 0,$$7-hour and $$R(s_t,g)=1 \;\; \text{if} \;\; S(s_t,g)\ge \delta_{\mathrm{done}} \;\; \text{else} \;\; 0,$$8-hour datasets: both are trained for $$R(s_t,g)=1 \;\; \text{if} \;\; S(s_t,g)\ge \delta_{\mathrm{done}} \;\; \text{else} \;\; 0,$$9 steps, with approximately $\delta_{\mathrm{done}}=0.8$0 minutes of policy optimization. Feature extraction, not RL optimization, becomes the dominant difference between the ViT-S and ViT-L variants (Liu et al., 27 Mar 2026).

The reported scaling figures are correspondingly simple: a $\delta_{\mathrm{done}}=0.8$1-hour dataset comprises approximately $\delta_{\mathrm{done}}=0.8$2 samples, a $\delta_{\mathrm{done}}=0.8$3-hour dataset approximately $\delta_{\mathrm{done}}=0.8$4 samples, ViT-L feature processing requires approximately $\delta_{\mathrm{done}}=0.8$5 minutes compared with approximately $\delta_{\mathrm{done}}=0.8$6 minutes for ViT-S, and the policy remains deployable after these stages without human intervention (Liu et al., 27 Mar 2026).

5. Empirical performance

Simulation experiments include a PointMaze-like environment with 3D ground-truth state. Coverage is evaluated with normalized entropies

$\delta_{\mathrm{done}}=0.8$7

applied to state entropy $\delta_{\mathrm{done}}=0.8$8, action entropy $\delta_{\mathrm{done}}=0.8$9, and joint entropy $S(s_t,g)=\frac{1}{4}\sum_{k=t-3}^{t}\cos(o'_k,g).$0. Under this metric, pink uniform exploration yields the highest coverage. In the same simulation study, the downstream policy success rate improves with data budget from $S(s_t,g)=\frac{1}{4}\sum_{k=t-3}^{t}\cos(o'_k,g).$1 to $S(s_t,g)=\frac{1}{4}\sum_{k=t-3}^{t}\cos(o'_k,g).$2 hours, and fitted Q-evaluation correlates strongly with true success rate, with Spearman $S(s_t,g)=\frac{1}{4}\sum_{k=t-3}^{t}\cos(o'_k,g).$3 (Liu et al., 27 Mar 2026).

Real-world evaluation uses a Unitree Go2 quadruped in three environments, denoted Simple, Standard, and Complex, with $S(s_t,g)=\frac{1}{4}\sum_{k=t-3}^{t}\cos(o'_k,g).$4 spatially diverse image goals and $S(s_t,g)=\frac{1}{4}\sum_{k=t-3}^{t}\cos(o'_k,g).$5 trials per goal, for maximum episode lengths of $S(s_t,g)=\frac{1}{4}\sum_{k=t-3}^{t}\cos(o'_k,g).$6 to $S(s_t,g)=\frac{1}{4}\sum_{k=t-3}^{t}\cos(o'_k,g).$7 steps. Reported baselines are Random, NoMaD, ViNT, and GNM. The principal metrics are success rate, completion time, and success weighted by time,

$S(s_t,g)=\frac{1}{4}\sum_{k=t-3}^{t}\cos(o'_k,g).$8

Among the highlighted results, MINav with a $S(s_t,g)=\frac{1}{4}\sum_{k=t-3}^{t}\cos(o'_k,g).$9-hour dataset and ViT-S exceeds all baselines on success rate in the Simple environment, achieving $\gamma=0.99$0 compared with a best baseline value of $\gamma=0.99$1. Increasing data from $\gamma=0.99$2 hour to $\gamma=0.99$3 hours raises success rate in the Standard environment from $\gamma=0.99$4 to $\gamma=0.99$5, and in the Complex environment from $\gamma=0.99$6 to $\gamma=0.99$7. ViT-L provides only slight gains, and only in the Complex environment at $\gamma=0.99$8 hours (Liu et al., 27 Mar 2026).

A real-world ablation on noise strategies, using $\gamma=0.99$9-hour datasets in the Standard environment, reports the following:

Noise strategy	SR	Time	STL
White Uniform	$H\to\infty$0	$H\to\infty$1	$H\to\infty$2
Pink Gauss	$H\to\infty$3	$H\to\infty$4	$H\to\infty$5
Pink Uniform	$H\to\infty$6	$H\to\infty$7	$H\to\infty$8

The reported robustness study uses the $H\to\infty$9-hour Standard-environment policy under dynamic human interference and on a second platform, the AgileX Limo. On the quadruped, the static and dynamic settings both achieve $o'_t=\phi(o_t)\in\mathbb{R}^d$00 success, with completion times of $o'_t=\phi(o_t)\in\mathbb{R}^d$01 and $o'_t=\phi(o_t)\in\mathbb{R}^d$02 and STL values of $o'_t=\phi(o_t)\in\mathbb{R}^d$03 and $o'_t=\phi(o_t)\in\mathbb{R}^d$04, respectively. On the wheeled platform, the static setting reaches $o'_t=\phi(o_t)\in\mathbb{R}^d$05 success with $o'_t=\phi(o_t)\in\mathbb{R}^d$06 completion time and $o'_t=\phi(o_t)\in\mathbb{R}^d$07 STL, while the dynamic setting reaches $o'_t=\phi(o_t)\in\mathbb{R}^d$08 success with $o'_t=\phi(o_t)\in\mathbb{R}^d$09 completion time and $o'_t=\phi(o_t)\in\mathbb{R}^d$10 STL (Liu et al., 27 Mar 2026).

Taken together, these results position MINav as an explicitly in-domain method rather than a general zero-shot navigator. Its strongest outcomes are reported after collecting data in the target environment and training an offline policy for that environment.

6. Scope, limitations, and related usage of the name

The limitations identified for MINav are concrete. Goal ambiguity can arise because near-wall observations may satisfy $o'_t=\phi(o_t)\in\mathbb{R}^d$11 in representation space without corresponding to true arrival; suggested remedies are stricter SSD filtering or depth. Partial observability and the absence of memory lead to failures on long chains of turns; adding recurrent or LSTM context is proposed as a possible improvement. The sparse reward may slow learning when the similarity margin is small; a learned dense reward via contrastive or regression objectives is suggested as a direction for improving sample efficiency. The policy is also described as blind to robot pose drift, which may be mitigated by fusing odometry or SLAM poses. Finally, the offline-only setting limits continual adaptation, and fitted Q-evaluation may inherit bias from the learned value function (Liu et al., 27 Mar 2026).

A recurrent source of confusion is nomenclature. In the same period, a distinct system named “MinNav” was introduced for tiny aerial robots, using a monocular camera, optical flow, uncertainty estimation, and active perception to navigate through static obstacles, dynamic obstacles, and unknown-shaped gaps (Patil et al., 5 Jun 2026). That method concerns active monocular navigation for aerial robots, whereas MINav concerns image-goal navigation via unsupervised exploration and offline RL for ground robots (Liu et al., 27 Mar 2026). The similarity of names does not indicate a shared methodology.

Within current visual navigation research, MINav is best understood as a compact pipeline defined by five coupled design choices: pink-uniform unsupervised exploration, frozen DINOv3 embeddings, hindsight goal relabeling, offline TD3+BC, and FQE-based checkpoint selection (Liu et al., 27 Mar 2026). The reported results suggest that these components, when assembled under tight runtime and hardware constraints, are sufficient for rapid policy prototyping and deployment in target environments.