bhatvineet/sharpa-dexmimicgen-upright-ppo

Sharpa-DexMimicGen: upright HOPE-can grasp residual PPO

Model repository: bhatvineet/sharpa-dexmimicgen-upright-ppo. Checkpoint redistribution terms have not yet been assigned a permissive license; public availability does not relicense HOPE objects, Wave meshes, or third-party assets. Verify their permissions before redistributing source assets or images.

Files and provenance

• residual_ppo.zip: final Stable-Baselines3 PPO checkpoint from 5,120 cumulative macro decisions (1,024 + 4,096 in separate runs).
• vecnormalize.pkl: required matching observation normalizer. Never load the policy with a different normalizer or with reward normalization enabled.
• SHA256SUMS: integrity fingerprints for checkpoint, normalizer and compact numerical demonstration ZIP. The checkpoint and normalizer are byte-identical to outputs/ppo_macro_pitch_expanded/last.zip and outputs/ppo_macro_pitch_expanded/vecnormalize_last.pkl in the source workspace.
• training_phase{1,2}_config.json: training settings and original run lineage.
• source_bundle.zip: small, separately documented Python source, conda environment and tests. No third-party meshes/textures or raw HDF5 are bundled.
• robot_action_pilot2.zip: ~1 MB, two accepted complete pick-to-bin numerical action/state traces (2 × 1,141 steps) and the five-attempt pilot manifest. No RGB images or original texture/mesh files are included. The full two RGB HDF5 demos are hosted separately in the private pickbin pilot dataset pending image redistribution review.

For a standalone checkout, download the full model repo into release/Sharpa-DexMimicGen/models/, then extract source_bundle.zip into release/ (the archive contains a top-level Sharpa-DexMimicGen/ directory). This places Sharpa-DexMimicGen/models/residual_ppo.zip alongside the Sharpa-DexMimicGen/sharpa_dexmimicgen/ package. Follow the extracted README.md to set external robot/object asset paths and create the conda environment; the checkpoint is not executable without those dependencies.

The model runs inside MacroResidualGraspEnv(repeat=10,pitch_prior=0.7) and requires a scripted template-53 Panda/Wave grasp to supply targets. It observes 125 normalized privileged simulated floats (including can pose/velocity, simulated fingertip contacts, robot state and nominal-target error) and emits 8 bounded wrist/finger residuals. It is not a camera model and does not control bin transport or release; the latter is scripted in the FSM.

Evidence and limitations

For hope_obj_001 on 100 held-out simulator resets with can yaw ±0.3 rad, XY sampler X∈[−0.13,−0.07], Y∈[−0.20,−0.14] m and fixed estimated dynamics, upright grasp/lift passed 98/100, compared with 70/100 for the matched +0.7 wrist-pitch scripted prior. Paired rescue/loss counts: 28/0. Acceptance required opposing pad contact, >10 cm lift, can axis ≤15° tilt, limited drift/penetration, sustained hold. Detailed per-seed results are in results/grasp_heldout_100.json on the proposed model hub and at ../results/grasp_heldout_100.json in this local code checkout. An isolated CPU conda re-evaluation on those identical 100 seeds (../results/conda_grasp_heldout_100.json) kept PPO at 98/100 but changed the scripted prior to 69/100 (one near-threshold case); do not conceal this runtime sensitivity.

This does not mean 98% real-robot or complete pick-and-place success. A separate new-scene FSM pilot passed 10/10 for the old narrow XY range, but only 21/30 with wider XY X∈[−0.18,−0.02], Y∈[−0.29,−0.05] m: 4 scripted setup collisions/disturbances and 5 PPO grasp failures. All 21 accepted grasps were scripted into the bin. Those are small, simulation-only pilots, not an independent 100-trial full-task study. Mass, friction, contact dynamics, hand mount, camera calibration and workspace coverage remain provisional. The PPO action-change smoothness statistic was slightly worse than the matched scripted prior despite better grasp yield.

Do not deserialize checkpoints from untrusted sources. SB3 ZIP / PyTorch serialization is executable-code-capable. Verify hashes and download only from a trusted release. External HOPE meshes and Wave robot descriptions are not included here and retain their own licenses.