Revised September 6, 2026. This article discusses simulation design decisions. The problem described in the earlier published B2-W record was that the robot stood but did not move forward. The changes included posture rewards, fall termination, separate leg/wheel evaluation and policy observations. Here we distinguish what each design changes. Robot and simulation footage (Japanese).
Separate staying upright from moving forward
The earlier article records a change from a body-height penalty to an upward-orientation reward. Maintaining posture alone does not establish velocity-command tracking. Investigating this problem calls for separate posture and speed-tracking reward terms alongside total reward and actual forward velocity. A higher reward must not automatically be labeled better walking.
Check fall termination together with rewards
The published record describes revising immediate termination on a fall to allow learning to continue afterward. Changing termination changes the states experienced during training. It does not establish that allowing falls improves walking.
A comparison should hold rewards and initial conditions fixed, change only termination, then inspect distance traveled, number of falls and recovery. If an episode continues after a fall, an unfinished episode must not be counted as a no-fall run.
Inspect individual leg and wheel motion
B2-W has 12 leg joints and four wheels. The record describes separating leg and wheel rewards, but we could not verify a controlled comparison isolating that change’s effect. Investigations should examine joint motion, wheel rotation and actual body displacement together. Spinning wheels without travel and travel driven only by the legs suggest different next checks. Joint-level records expose differences hidden by total reward.
Reducing observations is separate from physical deployment
The earlier article also describes a Blind Policy design that removes linear velocity from observations. This changes the information supplied to the policy. Fewer inputs do not establish robustness to real sensor errors or operation on a physical robot.
Physical integration needs checks of observation order, units and history, joint mapping and control period. This article does not claim verified physical-robot effects.
What the record supports
This is a review of changes documented in an earlier published record, not a controlled experiment isolating each change or a measured walking success rate. Design hypotheses must be distinguished from measured effects.
To evaluate the same problem next, retain a baseline and compare changes one at a time. Match velocity commands, terrain, initial states and trial durations; define success and failure beforehand. Logs and video from multiple trials then make those hypotheses testable.
LET’S BUILD TOGETHER
What are you working on?
Tell us about the software you need to implement or test. We can discuss the work, the development setup and how we could contribute.
Discuss your development needs