Ctrl-World RL4VLA Policy1000 Key-Moment Checkpoint Step 2000
This repository contains a Ctrl-World world-model checkpoint fine-tuned on the local RL4VLA / ManiSkill source pool.
File
checkpoint-2000.pt
Source Checkpoint
Local source path:
/root/workspace/CtrlWorld_RL4VLA_ArmSim/20260515_KeyMomentBalancedScale1000Sources_1000/results/model_ckpt/rl4vla_mainv3_policy1000_keymoment_balanced_step2000_noval_3gpu/checkpoint-2000.pt
SHA256:
e35d67127e3279fb6afaee4d6423fae05d8baa2967bfc05c1cb117889665502e
Training Data Summary
- Source pool:
rl4vla_mainv3_policy1000_keymoment_balanced - Environment:
PutOnPlateInScene25Main-v3 - Source policies:
sft,warmup_overlay - Action variants:
base,translation_scale2,translation_reverse,zero_translation - Train sources: 752
- Val sources: 248
- Train samples: 17,296
- Val samples: 5,704
- Slots per train source: 23
Training Notes
The run uses the Ctrl-World training entrypoint:
/root/workspace/0_Projects/Ctrl-World/scripts/train_wm.py
The checkpoint was initialized from the local pretrained Ctrl-World checkpoint:
/root/workspace/0_Models/hf_Ctrl-World/checkpoint-10000.pt
This is a diagnostic 2000-step checkpoint. Core training hyperparameters were kept from the Ctrl-World workflow; the run shortened max_train_steps and disabled validation-video export for checkpoint retention and downstream external evaluation.
Intended Use
This checkpoint was produced for short-horizon action-conditioned prediction and action-screening experiments in the RL4VLA / ManiSkill setting. It is not a general-purpose video generation model.