A counterexample shows how intermediate rewards can change the objective. Potential-based shaping, generated reward code and policy evaluation offer ways to check the effect.
Topic: #robotics
A collection of 10 posts about robotics.
Eureka, REvolve and RF-Agent introduce reward search. The lesson connects dopamine prediction errors, preference-based reward models, RLHF and RLAIF.
How does NVIDIA Flora record actions in robot video? The article compares model performance, throughput and cost under the reported conditions and examines failure detection.
The update sequence and residual equations show how tactile feedback corrects action chunks proposed by vision, alongside the experimental results and remaining questions.
The FoMo dataset includes raw multisensor records from 12 repeated seasonal runs over a year and 64 km. It lets researchers compare where SLAM fails as conditions change.
MIT’s SANDO plans paths through unknown environments using time-dependent corridors that account for moving obstacles. The analysis examines the continuous-time safety assumptions, the comparison behind a 7.4-fold optimization speedup and failures on hardware.
RAIBO2 combined mechanical design, motor drivers and a gait learned through reinforcement learning to finish a marathon. The race record also shows which aspects of a fully autonomous mission remain to be tested.
D4RT, the CVPR 2026 Best Paper, introduces 4D reconstruction through depth, tracking, camera pose, and query-based perception. Explore the evidence behind its results and the questions still open.
A review of “Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes.”
Data association and back-end optimization connect coordinate frames, uncertainty and loop closure in a SLAM system.