KND  a suite of the exam · Mobile robot

KinDER

Physical reasoning for planning and control: throw, balance, sweep, scoop, pack.

Physical reasoning for robot planning and control, kept apart from perception and language: a simulated TidyBot++ mobile manipulator throws, balances, sweeps, scoops, stacks into cupboards, sorts, clears and packs. Each episode passes or fails on KinDER's own goal check (an object's centre inside a goal region, and the like), with no partial credit. We keep 5 of its 31 task families (the 4 hardest, scored, and one fixture), one variant each, and run each in two modes: unlimited (full simulator state and the goal check, unlimited resets, one replayable trajectory handed in) and limited (a robot behind a service: its base and wrist cameras, odometry and joint readings; one episode without resets).

Originadapted from the upstream benchmark (5b2dbac (upstream), our selection of 4 tasks and a fixture (2026-10-04; 11 until then))
SimulatorMuJoCo 3.3 (KinDER's Dynamic3D environments)
RobotTidyBot++
Tasks5 tasks · 10 tests
Agent time30 min
01

Results

4 tasks
1 GPT-6.2 AstraCodex · high 50%2/4
2 Claude Fable 5.2Claude Code · high 25%1/4
2 GPT-6.1 SolCodex · high 25%1/4
4 Claude Opus 5.5Claude Code · high 0%0/4
4 DeepSeek V4.1 FlashDeepSeek Harness · high 0%0/4
4 Gemini 3.8 FlashAntigravity · high 0%0/4
4 Gemini 4 ArgonAntigravity · high 0%0/4
4 GLM 5.3 FlashClaude Code · high 0%0/4
4 GPT-6 AstraCodex · high 0%0/4
4 Kimi 3.1 ProClaude Code · high 0%0/4
02

Tasks

5 tasks · 10 tests