RoboDojo defines a standardized set of manipulation tasks run in simulation (Isaac Sim) and a corresponding set on physical robot hardware. Policies are run under a shared protocol and their results (task success) are aggregated into a single ranking on a public leaderboard. The RealEval layer standardizes real-world measurements to reduce hardware and procedural discrepancies across labs.
Evaluating robot manipulation policies has been fragmented: different works used incompatible tasks, simulators and protocols, preventing fair cross-model comparison and reliable sim-to-real transfer of results.
A set of manipulation tasks built in NVIDIA Isaac Sim, run under a shared protocol.
A set of tasks performed on physical robot hardware, verifying real manipulation ability.
A layer standardizing real-world measurements plus a public results leaderboard.
A separate, unrelated project RoboDojo.jl (a differentiable robotics simulator in Julia, based on the Dojo paper) exists — it is a different entity from this benchmark.
Real-world results depend on specific hardware and procedures — follow the protocol to keep comparability.
Release of the paper, leaderboard and reference implementation of the sim-and-real manipulation benchmark.
