Identity and contribution
Zeying Gong, Yangyi Zhong, Yiyi Ding, and colleagues introduce FLUX in arXiv:2603.12806v1, March 13, 2026; the project records IROS 2026. This navigation policy is unrelated to the image-generator model with the same name.
FLUX initializes from NavDP, replaces diffusion generation with Rectified Flow, then applies dynamic-environment RL. DynBench is its separate Isaac-based environment covering fixed/image goals, goal-free exploration, moving-person goals, crowd traversal, and dynamic exploration.
Generation and post-training
Eight RGB frames/current depth condition a sixteen-layer decoder and trajectory scorer. Sixteen generated candidates are selected by that scorer. A GRPO-style update without a separate value baseline does not mean the system has no candidate evaluator.
For Gaussian trajectory noise and expert path :
Euler integration generates the whole waypoint sequence. “Straight” means the training interpolation in trajectory space, not a straight physical route or an unconditional guarantee of straight learned integral curves.
The RF stage uses a 3,000-trajectory Gibson subset, thirty epochs, and a stated single RTX 4090. This inherits NavDP pretraining, so it is not a from-scratch total-cost comparison based only on the small subset.
Goal rewards combine +20 success, asymmetric progress weights, stagnation/timeouts, and social-distance penalties below 0.45/1.2 m. Exploration adds displacement and grid novelty. A novelty-reward narrative differs from its printed coefficients and needs code checking.
The paper standardizes discounted returns across timesteps within an episode, clips them to , and uses a clipped probability-ratio objective. This differs from classic same-state groups of independent full trajectories. The PDF does not fully specify flow-policy likelihood computation or how candidate selection maps to the update probability; the objective alone is insufficient for reproduction.
Crowd protocol and metrics
Training uses twenty clutter scenes and 20K episodes; evaluation uses six structured scenes and 600 episodes. SocialNav has ten to fifteen pedestrians, robot speed cap 0.5 m/s, and pedestrian speed 1.1 m/s. NavMesh/A* paths plus GoTo, Idle, and LookAround states define rule-based behavior, rather than learned social intent.
A robot moving less than 0.1 m over two seconds is stuck; timeout is 120 seconds; goal success is within one meter. Yielding pauses can be penalized by this protocol, and a one-meter moving-goal threshold differs from precise localization.
| Metric | NavDP | FLUX |
|---|---|---|
| Static PointNav SR / SPL | 77.8 / 74.8 | 80.9 / 78.6 |
| Dynamic PointNav success | 38.7% | 42.4% |
| Dynamic exploration time / area | 43.6 s / 93.7 m² | 55.1 s / 128.7 m² |
| SocialNav success | 59.3% | 64.0% |
| SocialNav Coll. / SC | 21 / 5.5% | 16 / 4.4% |
These NavDP values use this paper’s protocol, not NavDP’s original evaluation. SC is a lower-is-better fraction of time within 1.2 m, despite its “Social Compliance” label. Coll. is a below-0.45 m proximity event proxy, not measured physical-contact percentage.
The ImageNav table reports SR 44.0 and SPL 44.4. Standard same-episode SPL should not exceed SR, requiring an explanation. Claims of best performance on every metric also conflict with smaller collision/proximity values from some baselines.
Ablations, speed, and limitations
PointNav success progresses from deterministic regression 53.6%, DDPM 77.8%, conditional flow 78.9%, RF 80.4%, to RF+RL 80.9%. SocialNav similarly gives 49.3%, 59.3%, 60.6%, 62.8%, and 64.0%. Small static gains lack extensive uncertainty estimates; dynamic learning need not always improve static tasks.
Reported latency is 305.2 ms NavDP/ten steps, 408.6 ms FlowNav/ten, 244.9 ms FLUX/ten, and 216.5 ms FLUX/six. Approximately 29%/47% reductions use changed total sampling steps as well as models; they are not pure per-network-call gains.
Go2-W, Go2, and G1 demonstrations use a RealSense D455 and remote RTX 4060. They establish bounded deployment examples rather than equally sized real-robot comparison statistics.
Independent analysis: pair FLUX with a fixed-sensor/controller/timing benchmark, separately logging actual contacts, proximity duration, yielding, pedestrian response, and root/body behavior. DynBench’s state machine is not NavIsaacLab’s AMP full-body control. Bind method/environment commits, candidate counts, Euler steps, moving-goal update frequency, likelihoods, and metric code. Social intent and human comfort require additional evidence; the source notes did not execute the implementation.
Source references
The source versions and checks above are retained from the uploaded PaperRead notes. This English edition edits their presentation; it does not report new experiments.