Store
VisionLibra Robot Evals · Free Beta

Upload a robot log.
Get an eval report.

Task success, grasp success, recovery rate, human interventions per hour, and every failure with a mode and root cause — computed from your episode logs in minutes. No robot required to try it.

📥
Drop a robot log here, or click to choose
JSON / JSONL episode logs — analyzed instantly · ROS bag / MCAP — 30-second converter script · MP4 & RGB-D video scoring — private beta, email us
no log handy?

Parsed in your browser; only compact episode data (tasks, phases, timings — never raw video) is sent for analysis. Reports get a private share link. Free during beta.

Log Format

One flexible episode schema

Send a JSON array, {"episodes": [...]}, or JSONL (one episode per line). Field names are flexible — result/success/outcome, phases/timeline/events, duration or start/end all work. Failed episodes without a failure_mode are classified automatically from their phase timeline.

Coming from ROS? The converter script (pure Python, pip install rosbags, no ROS install) turns a ROS 1/2 bag into this schema in one command.

What you get

Report contents

  • Task success, grasp success and recovery-success rates
  • Human interventions per hour and median task time
  • Failure-mode distribution with per-episode root causes
  • Per-episode phase timelines, failures first
  • A private share link for your team
{ "episodes": [
  { "id": 1827,
    "task": "Pick bottle",
    "result": "fail",                    // or "success" (also accepts success: true/false)
    "duration": 9.0,                     // seconds — or provide start/end
    "phases": [                          // also accepts "timeline" / "events"
      { "t": 0,  "label": "approach" },
      { "t": 3,  "label": "pre-grasp" },
      { "t": 4,  "label": "grasp" },
      { "t": 5,  "label": "lift" },
      { "t": 6,  "label": "slip" },
      { "t": 7,  "label": "recovery attempt" },
      { "t": 9,  "label": "failure" } ],
    "failure_mode": "grasp_slip",        // optional — the AI labels missing ones
    "root_cause": "insufficient gripper closure",   // optional
    "recovery_attempted": true,
    "recovery_success": false,
    "interventions": 1                   // human interventions in this episode
  } ] }
Capture → Evaluate → Fail → Learn → Redeploy

Evaluation is step two of the loop

Robot Evals turns raw logs into the failure catalogue; VisionLibra Data Services and head-worn capture rigs generate the data that fixes what it finds. Policy-vs-policy regression comparison ships next — tell us what you need.