haldunbalim/GTPlanning

Generative Trajectory Planning (GTP): trained models

Trained GTP models from Model-Based Online Decision Making via Generative Trajectory Planning (NeurIPS 2026), for use with the code at github.com/haldunbalim/GTPlanning.

Paper · Code · Project page

Models

One model per environment, each with checkpoints every 100k environment steps from 100k to 1M.

Model	Environment	Return at 1M	Success at 1M
hopper	Hopper-v4	3556 ± 3
halfcheetah	HalfCheetah-v4	14885 ± 192
walker2d	Walker2d-v4	5403 ± 633
ant	Ant-v4	7138 ± 60
fetchpush	FetchPushDense-v4	-1.98 ± 0.90	100%
fetchpickandplace	FetchPickAndPlaceDense-v4	-2.45 ± 1.59	100%
fetchslide	FetchSlideDense-v4	-8.16 ± 3.87	72%
pusht	PushT-v0, from 96×96 images	-6.53 ± 5.35	100%

Usage

git clone https://github.com/haldunbalim/GTPlanning && cd GTPlanning
pip install -r requirements.txt && pip install -e .

python scripts/eval.py --model hopper                       # 1M checkpoint, 20 episodes
python scripts/eval.py --model fetchslide --step 500k 1M --episodes 50

The script downloads only the config and the requested checkpoints.

Files

<model>/config.yaml                       environment, network and planner settings, and the
                                          planning horizon of each checkpoint
<model>/checkpoints/step_XXXXXXX.msgpack  network weights and observation-normalization
                                          statistics (flax msgpack), every 100k steps

The planner's horizon grows over the first 300k (locomotion) or 350k (manipulation) training steps, so earlier checkpoints plan over a shorter horizon; config.yaml records it per checkpoint and scripts/eval.py applies it.

Citation

@inproceedings{balim2026modelbasedonline,
  title     = {Model-Based Online Decision Making via Generative Trajectory Planning},
  author    = {Balim, Haldun and Du, Yilun and Li, Na},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}