One policy must cover conflicting navigation demands
The paper frames local navigation as a partially observable decision problem: the robot receives LiDAR measurements, a relative goal, its motion state and its previous command, and must output body-frame longitudinal velocity, lateral velocity and yaw rate. Because motion and occlusion hide part of the scene, the policy conditions on a history of frames rather than a single snapshot.
The authors argue that a single policy struggles to serve scenarios with different decision demands at once: safe obstacle avoidance, narrow-passage traversal and escape from concave local traps. Their two stated questions are whether one policy can coordinate those skills from local observations alone, and whether that coordination survives unseen compositional scenarios and real deployment. The design choice follows from that framing: rather than adding explicit object tracking at deployment, they move privileged scene knowledge into training and compress it into a latent the deployed Student can reconstruct.
The Teacher sees obstacle tracks the Student never gets
TAME-Nav is an asymmetric Teacher–Student framework with separate encoders and a shared Actor. The Teacher combines clean observation histories with causal dynamic-obstacle tracks. Concretely, it encodes a ten-frame history into a 16-dimensional temporal latent, and separately encodes up to five nearest valid obstacles as 8-dimensional track features covering robot-relative position, body-frame obstacle velocity, circumscribed radius, an age-dependent uncertainty term, normalized observation age and visibility. Occluded tracks are extrapolated from their latest velocity estimates until they expire, and the paper states no future obstacle states are accessed.
Those track features pass through a shared MLP and self-attention, then are aggregated by cross-attention in which a robot-conditioned query from the current observation attends to the tracks. The resulting dynamic latent is fused with the temporal latent by normalized residual fusion, with a learnable scalar initialized to one. The Student, by contrast, sees only its own 55-dimensional observation — a 41-dimensional LiDAR scan plus 14 non-LiDAR quantities including the body-frame goal, velocities, previous command and projected gravity — and must infer a compatible latent from corrupted histories.
The shared Actor turns the latent into a nominal velocity mean, which a differentiable LiDAR-based safety correction adjusts before parameterizing a Gaussian policy. A privileged Critic, which also sees ground-truth dynamic-obstacle states and a static cost-to-go scalar, supports value estimation during training only. Deployment keeps the Student encoder, the shared Actor and the LiDAR safety module.
Soft routing gives the Student scene-dependent capacity
The Student's mixture of experts is soft rather than hard: a gating network maps the routing condition to expert weights through a softmax, and expert MLPs produce representations that are combined densely. The paper describes this architecture as providing a structural basis for learning scene-dependent representations across heterogeneous local-navigation situations, which is the mechanism by which one deployable policy is meant to cover the conflicting demands above.
Training splits rollouts across parallel environments, with 75% allocated to Teacher-controlled rollouts and 25% to Student-controlled rollouts. PPO updates the Teacher encoder, the shared Actor and the privileged Critic, while the Student encoder learns by latent distillation from stop-gradient Teacher representations at states the Student actually visits. That stop-gradient detail matters for reading the claim: the Student is pulled toward the Teacher's privileged latent, but the distillation target is not reshaped by Student gradients.
Coverage comes from procedural maps and asymmetric corruption
Training runs in a custom GPU-vectorized 2D simulator that updates poses and velocities through a command-response model instead of joint-level dynamics, while retaining footprint collisions with blocking and sliding, velocity and acceleration limits, dynamic obstacles, and LiDAR and odometry observations. A separately trained locomotion controller executes the same velocity commands at deployment, so the navigation policy is learned at the level of velocity commands rather than leg control.
The training set contains 600 layouts, evenly split across six scene families: dense clutter, multidirectional dynamic crossings, concave traps, irregular narrow corridors, occluded dynamic interactions and composite indoor layouts, with doorway traversal folded into indoor layouts. Procedural generation varies obstacle geometry, clearance, occlusion and dynamic motion, and each map supplies a prevalidated pool of start–goal configurations and initial headings. Scene labels, global maps and reference paths are excluded from the policy's inputs.
The asymmetry extends to randomization: the Student receives delayed and corrupted observations, while the Teacher and Critic retain clean observations of the underlying simulated state. The paper calls this deployment-informed asymmetric domain randomization and ties it to sim-to-real transfer, since sensor degradation and velocity-response uncertainty are injected on the Student side only.
Success gains exceed 25 points under matched budgets
Across in-distribution and unseen compositional scenarios, the paper reports that TAME-Nav achieves higher success rates, lower collision rates and higher navigation scores than representative baselines. The comparison is against SEA-Nav, NavRL and DRL-DCLP under matched training budgets, and the headline in-distribution figure is an improvement of over 25 percentage points in success over the strongest baseline. The advantage is reported to hold across unseen sequential and multiroom scenarios.
These are aggregate benchmark statements, and the packet does not break them down per scenario family or per baseline for every metric. Readers should therefore treat the success-rate gain, the collision-rate reduction and the navigation-score increase as three distinct reported comparisons rather than interchangeable evidence, and should not read the 25-point figure as applying to the unseen compositional settings, where the paper describes maintained advantage without a stated margin.
Real trials separate collision avoidance from trap escape
Real-world tests used a Unitree Go2 with a Livox MID-360 LiDAR and an onboard NVIDIA Jetson Orin, issuing velocity commands at 50 Hz while LiDAR updated at 10 Hz. In a local-trap scenario with a concave barrier blocking the direct route, each method ran ten trials, and a trial allowed one minor collision that did not topple an obstacle; a second collision, a toppling collision, or persistent entrapment without progress counted as failure. TAME-Nav completed all ten trials with only one minor collision. NavRL reached 50% success with 0.8 collisions per trial, with some escapes arising when large-amplitude oscillations reoriented the robot toward a traversable direction, and all remaining baselines failed.
The instructive contrast is SEA-Nav, which avoided collisions but stayed trapped. That result is the paper's clearest evidence that immediate collision avoidance and effective trap recovery are different capabilities, and that a policy can score well on safety while failing the task. The authors also note that the baselines trained on the same multi-scenario maps still did not achieve reliable trap escape in this real setting.
In a winding narrow passage where obstacles abruptly entered the path midway and near the goal, TAME-Nav completed all ten trials without collisions, while successful baselines incurred repeated contacts. Combined with the trap trials, that gives 20 of 20 real trials completed across the two scenarios with one minor contact in total. The scope is two constructed real-world scenarios on one quadruped platform, so the result supports robustness in those settings rather than a general claim about all deployment environments.
The privileged signal is the transferable asset
Read together, the results point to a specific priority rather than a generic endorsement of teacher-student training. The Teacher's advantage is not extra perception hardware at deployment but causal obstacle tracks and clean state during training; the Student's advantage is a temporal Soft-MoE encoder that can reconstruct a compatible latent from corrupted onboard histories. The 25-plus-point in-distribution success gain and the 20-of-20 real trials are the evidence that this particular privileged signal transfers into a deployable policy.
The boundary is equally specific. The benchmark gains are reported against three named baselines under matched budgets, and the real-robot evidence covers local-trap escape and narrow-passage navigation with sudden dynamic obstacles on a single quadruped. Within that scope, the paper's most decision-relevant distinction is between avoiding contact and completing the task: SEA-Nav's collision-free entrapment shows that a navigator judged only on collisions would look adequate while failing the scenario TAME-Nav was built to solve.
Sources used in this answer
TAME-Nav: Temporal Asymmetric Mixture-of-Experts Distillation for Agile and Versatile Robot Navigation
An asymmetric Teacher–Student framework distills track-augmented privileged representations into a temporal Soft-MoE Student that uses only onboard histories, improving success, collision rate and navigation score over SEA-Nav, NavRL and DRL-DCLP and completing 20 of 20 real quadruped trials with one minor contact.