REVERSAL-BENCH
REVERSAL-BENCH is an evaluation benchmark for reset-free reinforcement learning. It quantifies the degree of environment reversibility using a continuous reversibility parameter ρ, provides a reset oracle to verify state recoverability, and covers multiple physics engines and manipulation scenarios.
Timeline (last 90 days)
REVERSAL-BENCH reveals that the current mainstream reset-free reinforcement learning paradigm relies heavily on the reversibility of the underlying environment.
REVERSAL-BENCH introduces a continuous parameter ρ∈[0,1] to control the degree of environmental reversibility.
REVERSAL-BENCH provides a ground-truth reset oracle to determine whether a given state can truly be recovered.
REVERSAL-BENCH's reset oracle relies on underlying simulator state information (such as precise 3D object coordinates, velocity vectors, and contact constraints) to determine whether the current state is still within the basin of attraction of the reachable initial distribution.
REVERSAL-BENCH spans five physics engines and covers eight manipulation settings.
Many algorithms that perform well in reversible environments may suffer a cliff-like performance drop once they approach the irreversibility threshold.
All Facts (6)
REVERSAL-BENCH reveals that the current mainstream reset-free reinforcement learning paradigm relies heavily on the reversibility of the underlying environment.
50%UnverifiedREVERSAL-BENCH introduces a continuous parameter ρ∈[0,1] to control the degree of environmental reversibility.
50%UnverifiedREVERSAL-BENCH provides a ground-truth reset oracle to determine whether a given state can truly be recovered.
50%UnverifiedREVERSAL-BENCH's reset oracle relies on underlying simulator state information (such as precise 3D object coordinates, velocity vectors, and contact constraints) to determine whether the current state is still within the basin of attraction of the reachable initial distribution.
50%UnverifiedREVERSAL-BENCH spans five physics engines and covers eight manipulation settings.
50%UnverifiedMany algorithms that perform well in reversible environments may suffer a cliff-like performance drop once they approach the irreversibility threshold.
50%