Hopper Backflip Demonstration — extends past the hand engineerable reward of → Hopper (MuJoCo task)

The standard Hopper benchmark already has a reward that is easy to hand-engineer: the paper describes all of the MuJoCo tasks' reward functions, Hopper included, as simple quadratic functions of the robot's distances, positions, and velocities (deep-rl-human-prefs, §"3.1.1 Simulated Robotics", p. 6). The backflip demonstration reuses the identical simulated robot and action space but targets behaviors the paper itself frames, in its introduction, as ones for which "it is unclear how to hand-engineer a reward function that would incentivize them" (deep-rl-human-prefs, §"1 Introduction", p. 2). Because the robot's dynamics and observation space are unchanged between the two settings, the 900 queries needed to train the backflip in under an hour (deep-rl-human-prefs, §"3.2 Novel behaviors", p. 8) isolate what actually made the demonstration hard: not the physics, which the standard benchmark already solves cheaply, but the absence of any specifiable objective to hand-engineer in the first place.