Research map / Image generation / Fast diffusion sampling
Generative Trajectory Stitching through Diffusion Composition
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 CompDiffuser, a single bidirectional diffusion model trained on short overlapping trajectory chunks that stitches chunks at test time (autoregressively or in parallel) to solve long-horizon goal-reaching planning tasks.
Problem
Diffusion planners trained on trajectories fail to solve long-horizon tasks whose start-goal pairs differ from training data; trajectory stitching from short offline segments is needed without long-horizon training data.
Approach
Model the trajectory distribution as a product of conditional distributions over overlapping chunks, learned by one diffusion denoiser that conditions each chunk on its neighbors' noisy samples at the same noise level, enabling bidirectional information flow; at inference, chunks are composed with autoregressive (or parallel) denoising and blended.
Key results
- On OGBench PointMaze Giant stitch, CompDiffuser reaches 68±3% success vs. 29±3% for GSC and 0% for most offline RL baselines.
- On OGBench AntMaze Giant stitch, CompDiffuser achieves 65±3% success vs. 21% for HIQL and 20% for GSC.
- CompDiffuser resolves all PointMaze stitching tasks in the medium, large maze settings (100% success), outperforming baselines such as GSC at 93-99%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CompDiffuser instead of GSC | PointMaze stitch Giant · PointMaze stitch Giant | success rate: GSC 29 ± 3; CompDiffuser 68 ± 3 | Helped |
| Use CompDiffuser instead of HIQL | AntMaze stitch Giant · AntMaze stitch Giant | success rate: HIQL 21 ± 2; CompDiffuser 65 ± 3 | Helped |
| Use CompDiffuser instead of GSC | AntMaze explore Medium · AntMaze explore Medium | success rate: GSC 90 ± 2; CompDiffuser 81 ± 2 | Hurt |
| Use CompDiffuser instead of GSC | HumanoidMaze stitch Giant · HumanoidMaze stitch Giant | success rate: GSC 5 ± 1; CompDiffuser 67 ± 4 | Helped |
| Plan AntSoccer with 17D state rather than 4D state (ablation) | CompDiffuser · AntSoccer arena | success rate: 55 ± 6 -> 69 ± 3 | Helped |
| Plan AntSoccer with 17D state rather than 4D state (ablation) | CompDiffuser · AntSoccer medium | success rate: 13 ± 1 -> 17 ± 3 | Helped |
| Use autoregressive rather than parallel sampling (ablation) | CompDiffuser · Compositional planning tasks | plan quality | Helped |
| Use backward rather than forward autoregressive denoising order (ablation) | CompDiffuser · PointMaze stitch Medium, Large, and Giant | success rate: Medium: 100 ± 0 vs 100 ± 0; Large: 100 ± 0 vs 100 ± 0; Giant: 55 ± 4 vs 56 ± 2 | No clear effect |
| Vary the number of composed trajectories from 7 to 12 (ablation) | CompDiffuser · PointMaze stitch Giant | success rate: 51 ± 7; 53 ± 6; 55 ± 4; 56 ± 2; 50 ± 6; 50 ± 3 | No clear effect |
| Add replanning during rollout (ablation) | CompDiffuser · PointMaze and AntMaze stitch; Giant maze | success rate | Helped |
| Increase planning state dimension from 2D to 15D or 29D (ablation) | CompDiffuser · AntMaze Large and Giant | success rate | Hurt |
| Vary inverse dynamics model horizon (ablation) | CompDiffuser · AntMaze stitch Large and Giant | success rate: Large: 77, 86, 80, 85, 76, 77; Giant: 61, 65, 68, 63, 60, 63 | No clear effect |
Surprising findings
- Autoregressive sampling does not always help: parallel and AR sampling perform on par in the easier AntMaze Large setting, and forward vs. backward AR directions perform similarly.
Limitations
- Error accumulation over long chains of chunks can yield infeasible plans (e.g., passing through walls).
- Optimal number of test-time composed chunks K is task-dependent and must be tuned.
- Performance depends on a separately trained MLP inverse dynamics model, which can cause execution failures.
Experimental setup
Datasets: OGBench PointMaze stitch OGBench AntMaze stitch OGBench AntMaze explore OGBench HumanoidMaze stitch OGBench AntSoccer stitch PointMaze stitching datasets from prior work
Benchmarks: OGBench PointMaze OGBench AntMaze OGBench HumanoidMaze OGBench AntSoccer PointMaze stitching datasets (prior work)
Metrics: success rate
Baselines: Decision Diffuser Generative Skill Chaining RvS Decision Transformer GCBC GCIVL GCIQL QRL CRL HIQL
Builds on: Diffuser Decision Diffuser Generative Skill Chaining
Closest papers by meaning
- State-Covering Trajectory Stitching for Diffusion Planners (NeurIPS 2025 · similarity 0.93)
- Multi-Robot Motion Planning with Diffusion Models (ICLR 2025 · similarity 0.89)
- STITCH-OPE: Trajectory Stitching with Guided Diffusion for Off-Policy Evaluation (NeurIPS 2025 · similarity 0.88)
- Model-based Diffusion for Trajectory Optimization (NeurIPS 2024 · similarity 0.88)
- Coarse-to-Fine Compositional Diffusion for Long-Horizon Planning (NeurIPS 2026 · similarity 0.88)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.87)
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.