minphak/dexgraspvla-ckpt-check

DexGraspVLA — checkpoint evaluation evidence

Evaluation evidence for the DiT controller of a UR5e arm + Inspire RH56E2 hand in MuJoCo: the procedure, the exact tools, the raw outputs, the written reports, and the epoch54 checkpoint they were produced from.

This repo exists so that a number in a report can be traced back to the command that produced it. It is evidence, not a training set — there are no episodes here to train on. The training code lives in minphak/dexgraspvla-w2-code; the recorded episodes live in minphak/Dexgraspdatav2.

Read this before quoting any number

The headline result is 0 lifts in 5 episodes (50 mm criterion) for the epoch54 checkpoint, on the five objects it was trained on. That is a smoke test, not a success rate and not a test of generalization.

The perception half of that run is independently positive: the planner saw the real target in all five saved inputs, returned five object-appropriate boxes, and SAM+Cutie produced 0 empty mask chunks out of 50. So the no-lift result cannot be explained away by the planner being blind.

The head camera changed three times while these artifacts were produced, so this repo contains several evaluation generations that are not comparable with each other. Read docs/procedure/PROVENANCE.md before using any image as evidence. In short:

Outputs	Status
eval_results/01–08	Archived 2026-08-17 run on the nadir camera, table blown out by an unscoped headlight. Correct for what it was; does not show the current rig.
eval_results/10_rerun_20260819/	Current rig, but same invalid legacy planner crop as 11. Exposure images and tensor ablations valid; its 18 rollouts are not.
eval_results/11_epoch54_20260820/	Invalid legacy planner crop. Diagnostic only — do not quote its rollouts as policy performance.
eval_results/12_planner_check/	The corrected run. Dedicated nadir planner_cam, bbox reprojected into the oblique head_cam, SAM+Cutie. This is the one to cite.

Why 11 is invalid: the legacy PLANNER_CROP rectangle was calibrated for the old nadir head_cam. On the current oblique view it retains 0% of object pixels — it removed the target instead of isolating it, so the planner was handed empty background and invented an object.

Every grasp number in the dataset is scripted/expert replay. Everything in this repo is a learned flow-matching DiT policy. Always say which kind a number is.

Layout

Folders follow the project's main branch: written reports in docs/, scripts in tools/, rendered evaluation output in eval_results/. The numeric prefixes inside eval_results/ are kept, because they mark which evaluation generation a file belongs to (see the table above).

Path	Contents
docs/procedure/	CHECKLIST.md (the four checks and their pass criteria), PROVENANCE.md (what was generated when, on which camera), RIG_CAVEAT.md, RANDOMIZATION.md (why the table was pale)
tools/	The eval scripts that produced everything here, plus the shell drivers
eval_results/	Raw outputs: GIFs, PNGs, logs, JSON, CSV
docs/reports/	The written reports
checkpoint/epoch54.ckpt	The evaluated checkpoint (2.5 GB)
docs/CKPT_CHECK_README.md	The in-repo README this tree was copied from
docs/PROJECT_REPORT_3WEEKS.md	Full project narrative, 5–19 August

Start with docs/reports/EPOCH54_CORRECTED_PLANNER_EVAL_20260820.md.

Clone

Full clone, including the 2.5 GB checkpoint — needs git-lfs:

git lfs install
git clone https://huggingface.co/datasets/minphak/dexgraspvla-ckpt-check

Reports and logs only (skips the checkpoint and the ~274 MB of renders):

GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/minphak/dexgraspvla-ckpt-check

Or pull single files without cloning:

from huggingface_hub import hf_hub_download, snapshot_download

# one report
hf_hub_download("minphak/dexgraspvla-ckpt-check",
                "docs/reports/EPOCH54_CORRECTED_PLANNER_EVAL_20260820.md",
                repo_type="dataset")

# just the corrected run's evidence
snapshot_download("minphak/dexgraspvla-ckpt-check", repo_type="dataset",
                  allow_patterns=["eval_results/12_planner_check/*"])

# just the checkpoint
hf_hub_download("minphak/dexgraspvla-ckpt-check",
                "checkpoint/epoch54.ckpt", repo_type="dataset")

Using the checkpoint

checkpoint/epoch54.ckpt is a flow-matching DiT controller. It needs the training code from minphak/dexgraspvla-w2-code to load (0 missing / 0 unexpected keys); it has no noise_scheduler, by design.

Verified structure, read from the file itself:

Field	Value
Top-level keys	cfg, state_dicts, pickles
Model tensors	910
n_obs_steps	1
n_action_steps	64
policy._target_	controller.policy.dexgraspvla_controller.DexGraspVLAController
Size	2,525,496,419 bytes

Two things that will bite you:

• The epoch number is only in the filename. The checkpoint carries no epoch or global_step field — both read back None.
• cfg.policy.n_emb reads 768, but training overrides it to 1024. Trust the training config, not this field.

Training stopped at epoch 54 of a planned 130, so this is an underfit checkpoint. That is consistent with the 0-lift result and is the most likely explanation for it.

import torch
d = torch.load("epoch54.ckpt", map_location="cpu", weights_only=False)
model_sd = d["state_dicts"]["model"]   # 910 tensors

Reproducing an evaluation

The scripts in tools/ are copies of DexGraspVLA_Local/tools/poc/; they need the rig from the main project.

MUJOCO_GL=wgl PYTHONPATH=. python -u tools/poc/251_run_pipeline_e2e.py \
  --object 010 --episodes 4 --mask sam \
  --ckpt /path/to/epoch54.ckpt \
  --action-hold 150 --exec-steps 8 --max-steps 12000 \
  --z-probe --render --render-every 2 --out ./out

• MUJOCO_GL=wgl on Windows. egl is a hard crash there, not a slow fallback; egl is for the Linux container.
• --action-hold 150 matters. An earlier sweep reported 0/10 across all checkpoints purely because --action-hold 10 is too short for the arm to track the commanded pose (3.0 rad error). The predictions were fine.
• --mask sam needs the planner and an OpenRouter key.

Known-open issues

• 0 lifts on the checkpoint's own training objects (above).
• The policy ignores the mask channel: ablating the target mask moves the action 0.8% of its range, while right_state moves it 23%. It replays an average state-conditioned trajectory rather than a scene-conditioned one — so an approach that looks correct proves very little. See eval_results/06_ablation/.
• The training set is small and underfit: 386 episodes; 300 further recorded episodes were never added to the config.