Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
400
400
End of preview. Expand in Data Studio

CALVIN-MAX: human feasibility audit

A human audit of 400 cases from CALVIN-MAX, a benchmark of mid-execution changes on CALVIN's five-instruction chains. For each case, one annotator judged whether a competent robot could still complete the instruction from the state just after the change.

Result

No case was judged infeasible.

Change n Feasible Infeasible Unclear Feasible share (%), 95% Wilson interval
Target relocation 80 79 0 1 100 [95.4, 100]
Receptacle relocation 80 79 0 1 100 [95.4, 100]
Obstacle insertion 80 80 0 0 100 [95.4, 100]
Arm push 54 54 0 0 100 [93.4, 100]
Grasp loss 26 26 0 0 100 [87.1, 100]
Perceptual (five changes) 80 80 0 0 100 [95.4, 100]
All 400 398 0 2 100 [99.0, 100]

The share is computed over cases judged feasible or infeasible. The two unclear cases come from CALVIN's instructions not saying what contact between objects is allowed: a block moved deep onto the slider inside the cabinet (core-target_relocation-A-k3-E0247), and the drawer partly closed while a block was carried to it (core-receptacle_relocation-B-k4-E0106).

Cases

400 realized held-out cases from Max-A, using FLOWER's runs, drawn with a fixed seed (20261108). The draw favours changes that can make a task infeasible: 80 each for target relocation, receptacle relocation and obstacle insertion, 54 arm pushes, all 26 realized grasp losses, and 16 for each of the five perceptual changes (camera shift, sensor noise, illumination switch, visual-theme switch, distractor burst).

Each image shows the static camera (top) and the gripper camera (bottom), just before (left) and just after (right) the change. The frames were re-rendered from FLOWER's stored Base actions, with no policy in the loop. The annotator saw the instruction and a description of the change, never the rollout outcome.

Scope: the audit covers FLOWER's onset states, and only cases already accepted by the benchmark's automated checks.

Note: for one case (core-illumination_switch-A-k5-E0247, a lighting change during "open drawer"), a bug in the page shown during the audit left the instruction text empty; it was judged feasible from the frames and the change description. The released index.html and frames/tasks.json include the instruction.

Files

Path Content
index.html The labeling page. Open it in a browser next to frames/; answers are kept in the browser and exported as audit_answers.json.
frames/<case_id>.jpg Before and after frames for each case
frames/tasks.json The instruction (CALVIN subtask) of each case
audit_selection.json The 400 cases: change, variant, chain, instruction index, onset step, change parameters
audit_answers.json The annotator's answers (feasible, infeasible or unclear) with notes and timestamps
audit_summary.json Counts and Wilson intervals per change
audit_traces.tar.gz FLOWER's Base traces for the 236 chains, needed to re-render the frames

Reproduce

From the root of the CALVIN-MAX repository, with CALVIN installed and this dataset downloaded to <data>:

mkdir -p results/audit && cp <data>/audit_selection.json results/audit/
tar xzf <data>/audit_traces.tar.gz                         # -> results/base/flower_abc/traces/
python scripts/render_qual.py audit-render                 # before and after frames -> results/audit/frames/
python scripts/build_audit_page.py                         # -> results/audit/index.html
python scripts/analyze_audit.py results/audit/audit_answers.json

render_qual.py audit-select draws the cases again; it needs FLOWER's per-case Max-A results, which are not released.

To audit again with a new annotator, open index.html, answer every case and export.

Downloads last month
496