DemoGrasp-Force — Checkpoints & Training Runs

Pre-trained checkpoints and training logs for the DemoGrasp-Force project, which investigates whether force-aware training (reward shaping, observation augmentation, and VLM-predicted physical property conditioning) improves dexterous grasping under realistic physics.

Code: https://github.com/ruixucs/DemoGrasp-Force (commit 7f43ad6)

---

Repository Layout

.
├── ckpt/                             # 7 baseline checkpoints from the original DemoGrasp paper
│   ├── fr3_dclaw.pt
│   ├── fr3_panda_gripper.pt
│   ├── fr3_shadow.pt
│   ├── inspire.pt
│   ├── shadow.pt
│   ├── ur5_allegro.pt
│   └── ur5_svh.pt
└── runs_ppo/
    ├── E5_baseline/                  # E5 — binary reward, no force obs, no VLM (reference baseline)
    │   ├── config.json               # Full training config
    │   ├── events.out.tfevents...    # TensorBoard scalars
    │   └── model_{0..2100}.pt        # 22 checkpoints every 100 PPO steps
    └── E7_vlm/                       # E7 — binary reward, no force obs, + VLM physical-property conditioning
        ├── config.json
        ├── events.out.tfevents...
        └── model_{0..1400}.pt        # 15 checkpoints every 100 PPO steps

All checkpoints are raw PyTorch state_dict files saved via torch.save().

---

Experimental Setup

Item	Value
Hand	FR3 arm + Inspire dexterous hand (hand=fr3_inspire_tac)
Simulator	IsaacGym, 7000 parallel envs, RTX 3090
Objects	union_ycb_unidex/union_ycb_debugset.yaml
Physics	objectFriction=0.5 (paper uses 1.0 — more realistic / harder)
Algorithm	PPO (train=PPOOneStep), domain randomization on
Convergence criterion	Success-rate mean stable over 20+ readings

E5_baseline

Param	Value
task.env.rewardType	binary
task.env.observationType	eefpose+objinitpose+objpcl
Final success rate	~78% (mean over last 20 readings at step ~2100)
Convergence step	~1500–2000

E7_vlm

Param	Value
task.env.rewardType	binary
task.env.observationType	eefpose+objinitpose+objpcl+vlmprop
Final success rate	~73–75% (mean oscillating 72–75% from step 700–1431)
Peak mean	74.8 % at step 1353
Convergence step	~700–800 (plateau)
Finding	VLM physical-property conditioning alone yields a −3 to −5 % drop vs E5

See docs/experiment_results.md in the GitHub repo for the full 2×2×2 factorial design and ongoing E6, E8–E12 runs.

---

Usage

1. Download

With huggingface_hub:

from huggingface_hub import snapshot_download

# Everything
local = snapshot_download(repo_id="rayxu2005/DemoGrasp-Force")

# Just one file
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download(
    repo_id="rayxu2005/DemoGrasp-Force",
    filename="runs_ppo/E5_baseline/model_2100.pt",
)

Or via git:

git lfs install
git clone https://huggingface.co/rayxu2005/DemoGrasp-Force

2. Load a checkpoint

import torch
state = torch.load("runs_ppo/E5_baseline/model_2100.pt", map_location="cpu")
# state contains 'model' (policy weights) and PPO optimizer state

3. Run inference

Use the code from https://github.com/ruixucs/DemoGrasp-Force and point play_policy.sh at the downloaded checkpoint.

---

Notes & Caveats

• Failed runs excluded: E6, E8, E9, E10, E11, E12 were killed early (Ninja build error or OOM) and are not included. See the GitHub repo for the rerun script (scripts/run_experiments_e6.sh).
• Training speed: ~40 min per 100 PPO steps on RTX 3090, roughly 5.7× slower than the paper's RTX 4090 timings. A full 20 000-step training would take ~125 h.
• Lower absolute success rate: Our ~78 % (E5) vs the paper's ~96 % is explained by harder physics (objectFriction=0.5 vs 1.0) and a slower GPU. Relative comparisons across E5–E12 are what this repo is designed to support.

---

Citation

If you use these checkpoints, please also cite the original DemoGrasp paper (see the GitHub repo for details).