Sony

The 10 AI and machine learning patents reflect a broad push to automate content personalization and development workflows. One application scores games and media by cognitive load, combining asset variety and interaction patterns into user-specific complexity estimates that feed play-time suggestions and Hardware compatibility checks. Another translates a developer's plain-English request into the right asset from a studio repository, pairs it with usage instructions, and sequences delivery by available bandwidth. A third trains a model on recorded NPC encounters and uses the output to modify appearance, behavior, emotion, or role, either injected live or shipped in a new build. An AI teammate patent describes a system that accepts a high-level goal, watches how the player acts, and silently shifts its own actions to complement the player's style without giving away the win. A fifth filing generates game scenarios across Audio, video, and gameplay styles, then applies a kernel function over player behavior to build a profile that steers the game state toward what each person engages with. Sony also filed a patent that trains a model on highly viewed videos, scores draft game scenes and trailers for predicted popularity, and flags low-scoring elements before release. A seventh uses an LLM to read source code, pick where to insert API calls, and write them so an AI agent can extract the data it needs to play and test the game. An eighth describes a console system that recommends content with an explanation tied to scored user propensities, then retrains on whether the player acts on the suggestion. A ninth application maps plain-language wishes such as a more cinematic or faster feel into concrete console and game settings, pushing the configuration straight to the screen instead of requiring menu navigation. The final AI patent reads image frames from one play session, extracts context like bosses, settings, and item loadouts, and sends it through an API to an AI agent that plays and tests the game in a later session.

Four cloud gaming patents tackle bandwidth, fidelity, and data collection. One ranks game objects by priority, using targets, player gaze, and playable characters to assign higher frame rate and resolution to important elements while text and stats panels receive less, easing bandwidth and latency strain. A second pulls text draw calls out of the render pipeline and streams text separately from the encoded scene, so subtitles and UI stay readable when the connection degrades. A third taps cloud sessions to capture video, Audio, and controller input in sync, tagged by game ID and context like bosses or loadouts, without the player noticing, to feed debugging and agent support. The fourth encodes per-frame display metadata alongside the game stream so a player's client device can make display decisions from the rendered frames it receives.

Platform patents number 3 and span development environments, energy awareness, and session invitations. A cloud dev environment lets many game developers edit the same code in shared workspaces or isolate risky edits in cloned sandboxes that copy only the block being edited and link to unedited code and assets in the shared workspace, cutting rebuilds and bandwidth. Another filing has a console read grid supply and load data, then schedule downloads, throttle patch speeds, or lower frame rate and brightness when clean power is scarce or the grid is strained. A session share link lets a player send one URL that carries the game and session IDs, then drops the recipient into the same multiplayer session from whichever Platforms they pick, friend or not, and can be re-shared while keeping its metadata.

Two UI and UX applications explore voice control and non-verbal assistance. A voice control system tells players apart by voice and position, filters overlapping speech, and runs several players' spoken commands in parallel next to normal controller input, treating in-session chat as a command source alongside physical buttons. A game assistant patent lets players talk to an LLM-based helper and receive answers without words, using screen border colors, lit controller buttons, directional rumble, and yes/no sounds so voice help never talks over the game.

Graphics patents number 2 and both address fidelity allocation. An AI model watches player input and live game state to guess what an in-game asset is being used for, then renders and streams that asset at fidelity matched to that use instead of spending GPU cycles uniformly, regenerating the parameter from the same signals that reveal the asset's changing purpose. A second filing generates lightweight 3D placeholder representations of heavy game assets from virtual camera captures and hands them to the client rendering engine via metadata, cutting what a cloud gaming client must handle by combining placeholder generation and metadata-driven handoff in a single cloud-side pipeline.

Two VR and AR filings manage tracking handoffs and content remapping. When camera or IMU confidence drops, one system waits for the user's blink or saccade before switching tracking modes, so the visible jump in the VR view is hidden during an eye event reported by an eye tracking device. The other pulls 3D characters out of a running game, remaps them into a player's living room, and shows them on an AR headset, all without changing the underlying game program by intercepting models from the running title and re-rendering them in a different space.

A single Audio patent describes a system that turns a player's own in-game voice into a socialization profile. Mood, intensity, emotion, and personality inferred from voice output are written to a profile that drives player selection and Audio filtering in the same session, closing the loop between the voice a player produces and both their matchmaking pool and their live Audio feed through filtering, muting, volume regulation, and word replacement.

The Streaming category holds 1 patent on game-aware video encoding. The mechanism uses the game's own state, specifically the beginning or occurrence of a noninteractive sequence or cutscene, as the trigger to boost B-frames in a group of frames, cutting bandwidth when latency matters less while keeping interactive gameplay responsive.

A game mechanics patent fills virtual concert venues with NPC crowd members that copy the movements of real performers and audience members, so sparse metaverse events still feel lively. A majority/minority split lets most NPCs follow one movement pattern while a smaller group follows another, creating a crowd that is coordinated but not uniform.

One Hardware filing extends gameplay beyond the screen. An ML pipeline reads gameplay Audio, video, and haptics, then drives room lights, fans, blinds, and spare speakers, while skipping redundant effects and compensating for each device's latency using predicted future feature values, device position relative to the player, and a gameplay experience quality score that adjusts feature weights from user feedback.

The Networking category contains 1 patent on wireless interference prediction. Instead of reacting after a link has degraded, the gaming device predicts expected interference and adjusts the transmit power of specific hosts, selecting among several candidate combinations of hosts and power settings the one most likely to avoid interference while staying above a latency, packet loss, and bandwidth threshold, using gameplay-side feature data such as player movement, user input, frame rate, and computation load as inputs.