Reinforcement Learning News
The latest Reinforcement Learning coverage curated for game developers.
arXiv cs.GR details a voxel mapping system that uses semantic uncertainty, curvature, and texture to decide where to spend memory. The team also adds an RL controller that lets developers …
arXiv cs.GR details InterEvolve, a test-time method that evolves reward programs so humanoid controllers can tackle new loco-manipulation tasks without retraining. The system pairs an object-aware FB model with LLM-guided …
arXiv cs.GR details a design tool that can rapidly test whether 3D obstacle-course layouts are solvable under designer-set constraints. It pairs GPU-accelerated simulation with search and RL guidance, giving level …
ViCo is a new training framework aimed at making AI-generated charts look and read more like human-made figures. It combines self-reflection, multi-step reinforcement learning, and automated visual checks to improve …
A new muscle-driven simulation pipeline can generate realistic sprinting and drill motions without motion-capture demonstrations. Running at up to 1000x real time on a single GPU, it trains control policies …
A new muscle-driven locomotion system uses a fixed reflex controller plus reinforcement learning to tune just four biomechanical parameters. The result is more plausible gait, better symmetry, and stronger robustness …
InstantMimic pushes physics-based character training into the seconds range by moving the entire RL loop onto the GPU. The system targets imitation-driven motion control, cutting out CPU bottlenecks and fragmented …
A new musculoskeletal RL controller can produce human-like sprinting with only a minimal task reward and about an hour of training. The ?bb-hold approach cuts the action space down to …
RealMat combines Stable Diffusion XL with reinforcement learning to generate more believable material maps for 3D authoring. The pipeline starts from synthetic 2×2 material grids, then pushes the model toward …
RL-Lock brings reinforcement learning to interlocking assembly generation, turning voxel decomposition into a sequential decision problem. The system uses structured action chunking plus MCTS-guided policy-value learning to search huge combinatorial …
ThinkBLOX is a new VLM-driven pipeline for generating 3D indoor scenes through progressive reasoning instead of one-shot layout planning. It iteratively places and refines objects, aiming to reduce the awkward …
A SIGGRAPH 2026 paper reports a 99.98% motion-reproduction success rate using a tokenized, GPT-style controller trained on large motion datasets. For animation teams, the interesting part is the shift from …
Researchers trained a goal-conditioned PPO policy to estimate food material parameters from fracture behavior in a single forward pass, then used CMA-ES to refine the result. On orange peeling, the …
This paper shows a way to train one controller that can both mimic reference parkour motion and still adapt when the environment changes. For teams working on character movement, the …
TacCoRL reports a 72.5% average success rate on four contact-heavy bimanual tasks by adding tactile feedback to VLA policies and training in simulation. The practical takeaway is that the model …
SCRIPT pushes language-driven humanoid control toward more stable, physically plausible motion. The system combines a diffusion policy with multi-stage training, then adds reinforcement learning to improve instruction following and motion …
NaP-Control uses reinforcement learning to steer a diffusion policy’s latent noise, aiming for whole-body character control that is both fast and robust. The approach skips iterative test-time guidance, which can …
A new arXiv paper describes COSMO-Agent, a tool-augmented RL setup that lets an LLM drive a closed-loop CAD-to-simulation workflow. The interesting bit for game teams is the orchestration pattern: external …
AMD Schola v2.1 adds StateTree support and a more scalable training pipeline, which makes it more practical for teams using Unreal Engine to train and deploy agent behavior. The big …
A new framework for motion retargeting using reinforcement learning addresses common issues like foot sliding and self-collisions. This innovation is particularly relevant for programmers and animators working with robotics and …