Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Trajectory Diffusion for ObjectGoal Navigation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes trajectory diffusion (T-Diff), a DiT-based DDPM that generates future trajectory sequences conditioned on a semantic map and target object for ObjectNav, trained on FMM-collected optimal trajectories.
Problem
ObjectNav planners are single-step (end-to-end or modular), which ignores temporal consistency and leads to myopic planning; modular waypoint predictors also overfit to training room layouts.
Approach
Model the conditional distribution of future trajectory segments given the current semantic map and goal with a transformer diffusion model (DiT-style, DDPM), then drive a local policy toward the k_g-th waypoint of the generated sequence, replanning every 5 steps.
Key results
- On Gibson val, T-Diff reaches 79.6% SR and 44.9% SPL with DTS 1.00m, versus PONI at 73.6% SR and 41.0% SPL.
- On MP3D val, T-Diff reaches 39.6% SR and 15.2% SPL, exceeding the prior best SGM (37.7% SR, 14.7% SPL) in the reported table.
- Replacing the diffusion model with a same-size MSE-trained transformer decoder drops Gibson SR to 59.2% (MSE 0.6541), supporting the use of DDPM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use T-Diff instead of L3MVN | ObjectNav · Gibson val | SR (%): 76.9 -> 79.6 | Helped |
| use T-Diff instead of SGM | ObjectNav · Gibson val | SPL (%): 44.0 -> 44.9 | Helped |
| use T-Diff instead of L3MVN | ObjectNav · Gibson val | DTS (m): 1.01 -> 1.00 | Helped |
| use T-Diff instead of SGM | ObjectNav · MP3D val | SR (%): 37.7 -> 39.6 | Helped |
| use T-Diff instead of SGM | ObjectNav · MP3D val | SPL (%): 14.7 -> 15.2 | Helped |
| use T-Diff instead of Stubborn | ObjectNav · MP3D val | DTS (m): 5.01 -> 5.16 | Hurt |
| add sequence trajectory training (length 32) instead of a single waypoint (length 1, i-th = 1) (ablation) | T-Diff · Gibson val | SR (%): 71.0 -> 79.6 | Helped |
| use trajectory length 32 rather than length 4 (ablation) | T-Diff · Gibson val | SR (%): 75.9 -> 79.6 | Helped |
| condition trajectory generation on semantic map rather than RGB observation only (ablation) | T-Diff+LP · Gibson val | SR (%): 73.5 -> 76.9 | Helped |
| add the target object as a trajectory-generation condition (ablation) | T-Diff+LP · Gibson val | SR (%): 76.9 -> 79.6 | Helped |
| use trajectory diffusion rather than a simple MSE-trained decoder | trajectory generation · Gibson val | MSE: 0.6541 -> 0.0357 | Helped |
| use trajectory diffusion rather than a simple MSE-trained decoder | ObjectNav · Gibson val | SR (%): 59.2 -> 79.6 | Helped |
Surprising findings
- Predicting single waypoints (trajectory length 1) performed worse than predicting 32-step trajectories, indicating that sequence length matters for navigation accuracy.
- Training on MP3D and testing on Gibson gave T-Diff 78.2% SR while PONI dropped from 73.6% to 43.9% SR under the same cross-simulator setting.
Limitations
- Generated trajectories show prediction bias and use only the semantic map, history and target, not egocentric views or object relations.
- Training trajectories are mostly short and straight optimal paths, lacking complex multi-turn detour trajectories.
- No error bars are reported; the authors argue variance is small.
Experimental setup
Datasets: Gibson MP3D
Benchmarks: Gibson ObjectNav val MP3D ObjectNav val
Metrics: SR SPL DTS MSE
Baselines: DD-PPO EmbCLIP SemExp PONI OVG-Nav L3MVN SGM 3D-aware FBE ANS L2M Stubborn
Builds on: DDPM DiT PONI SemExp Fast Marching Method
Closest papers by meaning
- Versatile Navigation Under Partial Observability via Value-guided Diffusion Policy (CVPR 2024 · similarity 0.87)
- Navigation World Models (CVPR 2025 · similarity 0.87)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.87)
- Generative Trajectory Stitching through Diffusion Composition (NeurIPS 2025 · similarity 0.87)
- C-NAV: Towards Self-Evolving Continual Object Navigation in Open World (NeurIPS 2025 · similarity 0.87)
- Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.