The humanoid benchmark landscape.

27 whole-body humanoid tasks (locomotion + bimanual manipulation) on a simulated Unitree H1 with dexterous hands, built to stress hierarchical control and long-horizon reasoning.

An imitation-learning benchmark spanning humanoid, quadruped, and musculoskeletal embodiments with real motion-capture references, used to evaluate whole-body locomotion in MuJoCo.

A mobile bi-manual humanoid benchmark of 40 demonstration-driven household tasks requiring coordinated whole-body manipulation and locomotion.

A sim-to-real RL framework for humanoid locomotion with an Isaac Gym → MuJoCo verification stage, demonstrated zero-shot on physical RobotEra and Unitree hardware.

Learns a residual delta-action model from real rollouts to close the dynamics mismatch, enabling agile whole-body skills on a real Unitree G1.

A whole-body teleoperation and learning system for a Unitree H1 with dexterous hands, deploying sim-trained policies on real hardware across manipulation and locomotion.

A single neural whole-body controller distilled from a motion-imitation policy that unifies multiple control modes and transfers to a real Unitree H1.

A curriculum that learns fall-recovery / getting-up policies in simulation and deploys them on a real Unitree G1 — one of the few works targeting recovery, not nominal gait.

A multi-simulator (Isaac Gym / Isaac Sim / Genesis) RL framework for humanoid whole-body control — standardized training pipelines rather than a fixed task set or scoreboard.

An open-source, low-cost (~1 m, ~16 kg) humanoid with sim-trained locomotion deployed on accessible hardware, lowering the barrier to real-world humanoid experimentation.

A large-scale simulated kitchen suite (100 tasks across 120 scenes) for generalist manipulation and imitation learning, built on robosuite.

A 50-task robotic-arm manipulation benchmark for multi-task and meta-reinforcement learning — a long-standing standard for sim manipulation generalization.

A 100+ task vision-guided manipulation benchmark on a Franka arm in CoppeliaSim, widely used for language-conditioned and few-shot manipulation.

A benchmark for long-horizon, language-conditioned manipulation that scores chained instruction-following over sequences of tabletop tasks.

A GPU-accelerated (SAPIEN) benchmark and environment suite for generalizable manipulation, now in its third generation with large-scale parallel simulation.

A modular MuJoCo-based manipulation framework and standardized task suite that underpins many downstream benchmarks (including RoboCasa).

A lifelong-learning manipulation benchmark of 130 language-conditioned tasks designed to study knowledge transfer in robot policies.

A scalable dual-arm manipulation benchmark with strong domain randomization and automated data generation — the current iteration of the RoboTwin line.

A simulator built to predict real-robot manipulation performance — a cheaper, reproducible proxy that correlates with physical evaluations of generalist policies.

A high-performance embodied-AI simulator for navigation and rearrangement; Habitat 3.0 adds simulated humanoids for human-robot collaboration research.

An interactive 3D embodied-AI simulator and sim-to-real navigation benchmark for object navigation and task completion in household scenes.

1,000 everyday household activities in OmniGibson, targeting long-horizon, full-house embodied tasks aligned with human time-use surveys.

The standard benchmark family for instruction-following navigation, evaluating agents that follow natural-language route directions in photorealistic 3D scenes.

A physical manipulation benchmark with standardized 3D-printed objects and assembly tasks plus a large real-robot dataset, designed for reproducible hardware evaluation.

The de facto standard physical object set and protocols for benchmarking grasping and manipulation on real hardware across labs.

A large real-world grasping dataset and evaluation with a billion grasp-pose annotations across cluttered scenes — a reference benchmark for grasp detection.

Standardized physical task boards (peg-in-hole, connectors, fasteners) with timed scoring for repeatable, cross-lab assessment of real manipulation dexterity.

A recurring real-hardware competition at ICRA/IROS scoring grasping and manipulation on standardized tasks — one of the few sustained real-world manipulation contests.

An emerging class of distributed real-robot evaluation efforts (multi-lab comparison, autonomous scoring) seeking reproducible physical policy evaluation beyond simulation.

A consolidated cross-embodiment dataset (1M+ real-robot trajectories from 22 embodiments) that underpins generalist-policy training rather than a single scoreboard.

An industry-oriented effort to define repeatable real-world test protocols for humanoids in manufacturing — one of few initiatives aimed squarely at physical humanoid evaluation.

The landmark real-world humanoid disaster-response competition (driving, doors, valves, debris, stairs) that exposed how brittle physical humanoid autonomy was in the field.

A real-world teleoperated-avatar competition scoring remote manipulation, mobility, and human interaction through an embodied robot, judged on task and experience.

A large-scale real-world humanoid competition spanning athletics, combat, and task events — high-profile, but with limited standardized, reproducible scoring.

A 21 km real-world humanoid running event — a rare public endurance stress test, though battery swaps and low finish rates make it more spectacle than standardized benchmark.

Long-running real-world robot competitions: the Humanoid League scores bipedal soccer and the @Home League scores domestic service tasks and human interaction under standardized rules.

The core industrial-robot safety standards; ISO/TS 15066 defines collaborative force/pressure limits for human-robot contact — the basis for any physical safety evaluation.

The safety standard for personal-care robots (mobile servant, physical assistant, person carrier) — the closest existing standard to free-roaming humanoids around people.

An in-development standard explicitly scoping safety for industrial mobile robots including legged/humanoid form factors — the first standards-track work aimed at humanoids.

Standardized test methods for mobile-robot performance (navigation, docking, obstacle handling) and emergency-response robots — repeatable physical test artifacts and scoring.

Functional-safety and safety-sensor standards (SIL levels, protective equipment, smart sensing) governing the safety subsystems any deployed humanoid must satisfy.