ys1087@partner.kit.edu e7e3ae48f1 Add FastTD3 HoReKa experiment management system

- Fixed JAX/PyTorch dtype mismatch for successful training
- Added experiment plan with paper-accurate hyperparameters
- Created batch submission and monitoring scripts
- Cleaned up log files and updated gitignore
- Ready for systematic paper replication

2025-07-22 17:08:03 +02:00

1.8 KiB

Raw Blame History

FastTD3 HoReKa Experiment Plan

Added by Dominik - Paper Replication Study

✅ Proof of Concept Results

Initial Success: HoReKa Dev Run

Task: T1JoystickFlatTerrain
Duration: 7 minutes (5000 timesteps)
Performance: Successfully training at ~29 it/s
Key Achievement: Fixed JAX/PyTorch dtype mismatch issue (removed JAX_ENABLE_X64)
Status: ✅ Environment working, ready for full-scale experiments

Experiments to Replicate

Phase 1: MuJoCo Playground (Figure 11 from paper)

T1JoystickFlatTerrain (3600s)
T1JoystickRoughTerrain (3600s)
G1JoystickFlatTerrain (3600s)
G1JoystickRoughTerrain (3600s)

Hyperparameters (from paper):

total_timesteps: 500000
num_envs: 2048
batch_size: 32768
buffer_size: 102400 (50K per env)
eval_interval: 25000

Phase 2: IsaacLab (Figure 10 from paper)

Isaac-Velocity-Flat-G1-v0 (3600s)
Isaac-Velocity-Rough-G1-v0 (3600s)
Isaac-Repose-Cube-Allegro-Direct-v0 (3600s)
Isaac-Repose-Cube-Shadow-Direct-v0 (3600s)
Isaac-Velocity-Flat-H1-v0 (3600s)
Isaac-Velocity-Rough-H1-v0 (3600s)

Hyperparameters:

total_timesteps: 1000000
num_envs: 1024
batch_size: 32768
buffer_size: 51200
eval_interval: 50000

Phase 3: HumanoidBench (Figure 9 from paper - subset)

h1hand-walk (10800s)
h1hand-run (10800s)
h1hand-hurdle (10800s)
h1hand-stair (10800s)
h1hand-slide (10800s)

Hyperparameters:

total_timesteps: 2000000
num_envs: 256
batch_size: 16384
buffer_size: 12800
eval_interval: 100000

Usage

Submit Phase 1:

python submit_experiment_batch.py --phase 1 --seeds 3

Monitor progress:

python monitor_experiments.py --watch

1.8 KiB Raw Blame History