TontonTremblay/tiptop-robolab120

TiPToP on RoboLab-120 — Full planning and replay outputs

End-to-end evaluation of TiPToP on the 120-task RoboLab benchmark.

TiPToP consumes only RGB + depth + camera calibration from the simulator — no ground-truth world state is read. Perception uses the canonical TiPToP stack (Gemini 2.5 Flash + SAM2 + M2T2 + cuTAMP).

Headline result

Metric	Count	Rate
Planning success (cuTAMP)	52 / 120	43.3%
Replay success given a plan	22 / 52	42.3%
End-to-end success	22 / 120	18.3%

Setup used

• Camera: external_cam (third-person RGB-D, 1280×720).
• Grasp TCP: +1.5 cm deeper (TIPTOP_GRASP_DEEPER_M=0.015). TiPToP's own top-scoring M2T2 grasp selection is unchanged.
• TiPToP upstream from commit 9c42c46 (tiptop-robot/tiptop@main) with one upstream patch: null-guard in tiptop/perception/segmentation.py :: segment_pointcloud_by_masks (skip objects with <10 points above the table plane). No other benchmark-specific changes.
• Gemini model: gemini-2.5-flash.
• cuTAMP: max-planning-time=20 s, num-particles=64, opt-steps-per-skeleton=100.
• Replay: stride=4, max-joint-step=0.03 rad, gripper-steps=60, post-steps=120. Timeout 900 s per task (1800 s for retries).

See the TiPToP fork's robolab/docs/robolab.md for the full pipeline and install instructions: https://github.com/TontonTremblay/tiptop/tree/robolab120/robolab

Contents

README.md                             ← this file
TIPTOP_ROBOLAB120_REPORT.md           ← per-task Markdown table
h5/manifest.json                      ← RoboLab→TiPToP H5 export manifest (120 tasks)
planning_outputs/
├── summary.json                      ← rolling manifest (120 rows)
├── metrics.json                      ← aggregate planning metrics
└── <TaskName>/<YYYY-MM-DD_HH-MM-SS>/
    ├── metadata.json                 ← planning metadata
    ├── tiptop_plan.json              ← TiPToP plan (arm traj + gripper)
    ├── tiptop_run.log
    ├── tiptop.yml
    ├── bboxes_viz.png / masks_viz.png / rgb.png
    ├── cutamp/                       ← cuTAMP env/state snapshots
    └── perception/
        ├── bboxes.json               ← Gemini detections
        ├── masks.npz                 ← SAM2 masks
        ├── depth.png
        ├── intrinsics.json
        ├── pointcloud.ply            ← segmented scene point cloud
        ├── grasps.pt                 ← M2T2 grasp candidates
        └── cutamp_env.pkl
replay/
├── replay_metrics.json               ← combined replay metrics
├── replay_metrics.<worker>.json      ← per-worker metrics
├── replay_summary.<worker>.json(l)   ← per-worker per-task result stream
└── <TaskName>/
    ├── result.json                   ← RoboLab pass/fail + env_results
    ├── replay.mp4                    ← external + wrist frames (51 videos)
    ├── stdout.txt / stderr.txt

Failure breakdown

Of the 68 tasks that did not produce a plan:

• ~34 "No satisfying particles" (cuTAMP),
• ~13 "No valid plan skeletons",
• ~6 motion-planning failed,
• ~8 perception edge cases,
• ~7 no grounded atoms / other.

Of the 30 plans that did not succeed at replay:

• most are plans executed but RoboLab predicate false at t ≤ 900 — typical for multi-step tasks where TiPToP is open-loop and errors compound,
• a few replays hit the Isaac-sim wall-clock budget on very heavy scenes.

How to reproduce

See https://github.com/TontonTremblay/tiptop/tree/robolab120 and the top-level README "RoboLab-120 integration" section.

# 1. Export H5 observations (RGB+depth only).
$ROBOLAB_PYTHON robolab/scripts/export_robolab_tiptop_h5_rgbd.py \
    --headless --task $(cat robolab/robolab120_task_names.txt) \
    --output-dir tiptop_robolab120_external_h5 --camera external_cam

# 2. Plan.
python3 robolab/scripts/run_tiptop_robolab120_h5_deeper.py \
    --manifest tiptop_robolab120_external_h5/manifest.json \
    --output-dir tiptop_robolab120_external_deeper_outputs \
    --timeout-s 900 --max-planning-time 20 --num-particles 64 \
    --opt-steps-per-skeleton 100 --grasp-deeper-m 0.015 --mem-cap-gb 24

# 3. Replay in RoboLab.
python3 robolab/scripts/replay_tiptop_robolab_manifest.py \
    --planning-summary tiptop_robolab120_external_deeper_outputs/summary.json \
    --output-dir tiptop_robolab120_external_deeper_replay \
    --timeout-s 900 --stride 4 --max-joint-step 0.03 \
    --gripper-steps 60 --post-steps 120 --camera both \
    --robolab-python "$ROBOLAB_PYTHON"

License

• This dataset: MIT.
• TiPToP: MIT (upstream).
• RoboLab-120 tasks / renders: see the robolab_valts repo for their license.

Citation

@article{shen2026tiptop,
  title={{TiPToP}: A Modular Open-Vocabulary Planning System for Robotic Manipulation},
  author={Shen, William and Kumar, Nishanth and others},
  journal={arXiv preprint arXiv:2603.09971},
  year={2026}
}