iKCE diagnostic — walker-walk DreamerV3 artifacts
Trained checkpoints and saved rollouts that back the workshop paper "Imagined Rollouts Are Kinematic, Not Dynamic: A Diagnosis of Long-Horizon World-Model Failure" (Robot World Model Workshop @ RSS 2026).
This repo is consumed by the code at
github.com/TUM-AVS/iKCE;
see docs/CHECKPOINTS.md there for the full provenance and the
training commands. Fetch into a clone via the included
scripts/fetch_artifacts.sh, or manually with
hf download <this-repo> --local-dir . from the repo root.
Contents
checkpoints/
├── walker_walk/ # default checkpoint (seed 0, imag_horizon=15)
├── walker_walk_h64/ # actor-training-horizon ablation (imag_horizon=64)
├── walker_walk_seed1/ # seed-variability replicate (seed 1)
├── walker_walk_seed2/ # seed-variability replicate (seed 2)
└── walker_walk_dr/ # domain-randomization control
# (μ ∼ U(0.1, 1.7) per-episode at training time)
results/
├── walker_walk_wm_h64/ # WM-imagined rollouts under default actor
├── walker_walk_physics_policy_h64/ # physics rollouts under default actor
├── walker_walk_wm_h64_dr/ # WM-imagined rollouts under DR policy
└── walker_walk_physics_policy_h64_dr/ # physics rollouts under DR policy
Each checkpoints/<run>/ ships:
latest.pt— DreamerV3 weights (~219 MB), SHA-256 indocs/CHECKPOINTS.mdconfig.yaml— resolved config the iKCE adapter loads at inference timemetrics.jsonl— training metrics (per-evaltrain_return/eval_return)
Each results/<run>/<axis>/ ships:
rollout_NNN.pt— saved per-rollout state trajectories (friction: 20 rollouts per μ × 13 cells = 260;joint_noise: 20 × 10 cells = 200 per run-axis pair)ikce_summary.csv— bootstrap-aggregated iKCE per cell
License
MIT. See the LICENSE in the code repository.
Citation
If you use these artifacts, please cite the workshop paper. A machine-readable
CITATION.cff is included in this repository — Hugging Face renders it as the
"Cite this repository" box above — and also lives in the
code repository.