[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-pac-man-humanoid-dodgeball-safety-en":3,"article-related-pac-man-humanoid-dodgeball-safety-en":29,"series-research-00655d82-4035-4ac8-8c48-ed070e82f4a5":74},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":11},"00655d82-4035-4ac8-8c48-ed070e82f4a5","pac-man-humanoid-dodgeball-safety-en","PAC-MAN makes humanoid dodgeball safer","\u003Cp>What does PAC-MAN do for humanoid dodgeball?\u003C\u002Fp>\u003Cp data-speakable=\"summary\">PAC-MAN combines control-barrier safety with onboard perception to help a humanoid dodge balls in the real world.\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>Research org\u003C\u002Fstrong>: Unspecified in arXiv abstract\u003C\u002Fli>\u003Cli>\u003Cstrong>Core data\u003C\u002Fstrong>: 95% of throws\u003C\u002Fli>\u003Cli>\u003Cstrong>Breakthrough\u003C\u002Fstrong>: Perception-aware CBF-RL with segmentation-masked depth and link-level safety guidance\u003C\u002Fli>\u003C\u002Ful>\u003Cp>PAC-MAN is interesting because it tackles a problem robotics teams hit immediately in the real world: safety logic often assumes cleaner perception than a deployed robot actually gets. Here, the robot is not given a privileged, perfect view of the ball at runtime. Instead, it has to react from a head-mounted camera that only sees the ball as segmentation-masked depth, while still keeping its whole body out of trouble.\u003C\u002Fp>\u003Cp>That makes this paper relevant beyond dodgeball. Any humanoid system that has to move quickly around obstacles, people, tools, or moving objects faces the same tension between what is safe in simulation or training and what is observable on hardware. The paper’s main contribution is a way to keep the safety structure useful even when perception is imperfect.\u003C\u002Fp>\u003Ch2>What problem this paper is trying to fix\u003C\u002Fh2>\u003Cp>Humanoid robots are hard to protect because their bodies have many links, and a collision can happen anywhere, not just at the hands or feet. In a fast game like dodgeball, the robot also has to react to a moving object that may only be partially visible, briefly occluded, or hard to localize precisely from a single onboard camera.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785481389808-eysc.png\" alt=\"PAC-MAN makes humanoid dodgeball safer\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>The abstract frames the core issue clearly: control-barrier-function methods can provide safety guidance, but the value of that guidance depends on what the robot can actually observe. If the state estimate is too good in training and too weak at deployment, the policy may look strong on paper and then fall apart on hardware. PAC-MAN is built to narrow that gap.\u003C\u002Fp>\u003Cp>In other words, this is not just a “learn to dodge” paper. It is a “learn to dodge while respecting whole-body safety under realistic sensing” paper. That distinction matters for engineers, because deployment failures usually come from the mismatch between the training setup and the sensor package on the robot.\u003C\u002Fp>\u003Ch2>How the method works in plain English\u003C\u002Fh2>\u003Cp>PAC-MAN stands for Perception-Aware CBF-RL. The framework couples control-barrier-function safety with \u003Ca href=\"\u002Ftag\u002Freinforcement-learning\">reinforcement learning\u003C\u002Fa>, but it does so with a perception pipeline that matches deployment reality. At runtime, the policy sees the ball through a head-mounted camera, and the ball is represented as segmentation-masked depth rather than as a perfect state vector.\u003C\u002Fp>\u003Cp>During training, the system uses CBF guidance that represents clearance to every body link. That means the learning signal is not just about avoiding a single point of contact; it is about keeping the whole robot’s body safe. The abstract also says the method uses an adversarial motion prior to regularize the evasive reflexes, which suggests the policy is pushed toward more robust dodge behaviors rather than brittle, overfit motions.\u003C\u002Fp>\u003Cp>The paper compares different ways of using barrier structure. One of the key findings is that Joint-CBF works best when the ball state is accurate, but its performance drops when the robot only has fixed-camera observations and the barrier is used only as training guidance. It improves again if the system \u003Ca href=\"\u002Fnews\u002Fblack-duck-coverity-ai-era-triage-en\">gets better\u003C\u002Fa> ball tracking, either through a gimbal or a privileged runtime filter.\u003C\u002Fp>\u003Cp>That is an important practical point: barrier methods are not magic. They are only as effective as the observability of the state they depend on. PAC-MAN’s design acknowledges that constraint instead of pretending it away.\u003C\u002Fp>\u003Ch2>What the paper actually shows\u003C\u002Fh2>\u003Cp>The evaluation uses a controlled any-link contact \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> with seeded throws in two regimes: single throws, and a deployment loop where the robot walks back to its station and recovers between throws. That second setup matters because it tests whether the policy can keep working across repeated interactions instead of only on a one-off dodge.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785481381393-slrs.png\" alt=\"PAC-MAN makes humanoid dodgeball safer\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>The abstract does not give a full table of benchmark numbers, so there is no detailed score breakdown to cite here. What it does say is that the policy comes within a few points of a privileged state oracle, and that a fixed onboard camera alone is adequate for evasion. Those are strong claims, but they are still described at a high level in the abstract rather than with exact metrics.\u003C\u002Fp>\u003Cp>For real-world deployment, the authors say they use a lightweight Link-CBF policy zero-shot on the Unitree G1. In that setting, the robot tolerates imperfect perception, succeeds on 95% of throws, and uses semantic segmentation to dodge different balls. The “zero-shot” detail is especially relevant for practitioners: the policy is deployed without additional on-robot retraining in the abstract’s description.\u003C\u002Fp>\u003Cp>The paper also highlights a nuanced result about observability. Joint-CBF gives the best performance with accurate ball states, but degrades under fixed-camera observations when used only as training guidance. With a ball-tracking gimbal or a privileged runtime filter, the performance recovers. So the method is not one universal recipe; it is a framework whose best configuration depends on how much perception the robot can really trust.\u003C\u002Fp>\u003Ch2>Why developers should care\u003C\u002Fh2>\u003Cp>For robotics developers, the takeaway is straightforward: safety and perception need to be designed together. A beautiful controller that assumes perfect state estimates is not enough if the hardware only has a noisy camera and a limited field of view. PAC-MAN is a concrete example of how to align the safety mechanism with the sensor reality of deployment.\u003C\u002Fp>\u003Cp>The paper is also useful as a design pattern. It separates what is available at training time from what is available at runtime, then uses that mismatch deliberately instead of accidentally. That can be a helpful mental model for anyone building humanoids, mobile manipulators, or other robots that must react quickly under partial observability.\u003C\u002Fp>\u003Cp>There are still limitations. The abstract describes a controlled benchmark and a specific real-world dodgeball setup, so it does not prove general safety across arbitrary tasks, environments, or adversaries. It also does not provide the full benchmark numbers in the abstract, which means readers should treat the reported “within a few points” comparison as directional until they read the paper itself.\u003C\u002Fp>\u003Cp>Even so, the engineering signal is strong. If you are building embodied systems, the lesson is that perception-aware safety can be more practical than either pure model-based safety or pure learned reflexes on their own. PAC-MAN’s contribution is showing that a humanoid can dodge effectively with a realistic sensor stack while still using whole-body safety structure to guide behavior.\u003C\u002Fp>\u003Ch2>Bottom line\u003C\u002Fh2>\u003Cp>PAC-MAN is a robotics paper about making safety work under real sensing limits, not just in clean simulation. Its core idea is to fuse control-barrier-function guidance with onboard perception so a humanoid can dodge balls using the camera it actually has, not the perfect state it wishes it had.\u003C\u002Fp>\u003Cp>For engineers, that makes it a useful reference point for deploying fast-reacting policies on humanoids: keep the safety logic tied to the body, keep the perception assumptions honest, and expect performance to depend heavily on how observable the world really is.\u003C\u002Fp>\u003Cul>\u003Cli>Whole-body safety has to account for every link, not just a single contact point.\u003C\u002Fli>\u003Cli>Barrier methods work best when the runtime state is observable enough to support them.\u003C\u002Fli>\u003Cli>Deployment realism matters: fixed cameras, segmentation, and zero-shot transfer change the game.\u003C\u002Fli>\u003C\u002Ful>","PAC-MAN combines control-barrier safety with onboard perception to help a humanoid dodge balls in the real world.","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.28623",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785481389808-eysc.png","research","en","27ea61f5-fd76-4df9-a9d0-05f049ca9a66",[17,18,19,20,21],"humanoid robotics","control barrier functions","reinforcement learning","perception","safety",[23,24,25],"PAC-MAN combines CBF-guided learning with deployment-realistic onboard perception.","The paper shows observability strongly affects how well barrier-based safety works.","A zero-shot Link-CBF policy on Unitree G1 reportedly succeeds on 95% of throws.",2,"2026-07-31T07:02:32.928022+00:00","2026-07-31T07:02:32.922+00:00",{"tags":30,"relatedLang":33,"relatedPosts":37},[31],{"name":19,"slug":32},"reinforcement-learning",{"id":15,"slug":34,"title":35,"language":36},"pac-man-humanoid-dodgeball-safety-zh","PAC-MAN 讓人形機器人閃球更安全","zh",[38,44,50,56,62,68],{"id":39,"slug":40,"title":41,"cover_image":42,"image_url":42,"created_at":43,"category":13},"2844c399-e2ad-4cf0-b888-fd480d6d56ed","retoken-one-token-visual-retrieval-en","ReToken uses one learned token to fetch visual context","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785479567286-85y5.png","2026-07-31T06:32:24.643191+00:00",{"id":45,"slug":46,"title":47,"cover_image":48,"image_url":48,"created_at":49,"category":13},"660b1f2e-8988-4d2e-9429-b0c7accd3b6a","ml-trace-seiberg-dualities-en","ML helps trace Seiberg dualities","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785477772867-xtol.png","2026-07-31T06:02:28.714509+00:00",{"id":51,"slug":52,"title":53,"cover_image":54,"image_url":54,"created_at":55,"category":13},"8e53848b-d163-4335-a8e4-29694e85bbb3","fruitfly-inspired-regression-without-heavy-models-en","Fruitfly-Inspired Regression Without Heavy Models","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785394974203-76p6.png","2026-07-30T07:02:27.641563+00:00",{"id":57,"slug":58,"title":59,"cover_image":60,"image_url":60,"created_at":61,"category":13},"2515b20a-f125-4354-b386-50e75eff70c4","mental-world-modeling-simulating-minds-en","Mental World Modeling: Simulating minds, not just scenes","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785393182740-cnwi.png","2026-07-30T06:32:29.590725+00:00",{"id":63,"slug":64,"title":65,"cover_image":66,"image_url":66,"created_at":67,"category":13},"f234ef6f-2934-4e01-bc50-3132313c0d7a","pretrain-q-functions-online-rl-finetuning-en","Do You Need to Pretrain Q-Functions?","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785391372788-de5m.png","2026-07-30T06:02:24.317984+00:00",{"id":69,"slug":70,"title":71,"cover_image":72,"image_url":72,"created_at":73,"category":13},"654f2009-0838-4f91-946b-61e508f5ba9b","openai-agent-hack-forces-tighter-eval-controls-en","OpenAI’s agent hack forces tighter eval controls","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785326591602-45xj.png","2026-07-29T12:02:46.361703+00:00",[75,80,85,90,95,100,105,110,115,120],{"id":76,"slug":77,"title":78,"created_at":79},"a2715e72-1fe8-41b3-abb1-d0cf1f710189","ai-predictions-2026-big-changes-en","AI Predictions for 2026: Brace for Big Changes","2026-03-26T01:25:07.788356+00:00",{"id":81,"slug":82,"title":83,"created_at":84},"8404bd7b-4c2f-4109-9ec4-baf29d88af2b","ml-papers-of-the-week-github-research-desk-en","ML Papers of the Week Turns GitHub Into a Research Desk","2026-03-27T01:11:39.480259+00:00",{"id":86,"slug":87,"title":88,"created_at":89},"87897a94-8065-4464-a016-1f23e89e17cc","ai-ml-conferences-to-watch-in-2026-en","AI\u002FML Conferences to Watch in 2026","2026-03-27T01:51:54.184108+00:00",{"id":91,"slug":92,"title":93,"created_at":94},"6f1987cf-25f3-47a4-b3e6-db0997695be8","openclaw-agents-manipulated-self-sabotage-en","OpenClaw Agents Can Be Manipulated Into Failure","2026-03-28T03:03:18.899465+00:00",{"id":96,"slug":97,"title":98,"created_at":99},"a53571ad-735a-4178-9f93-cb09b699d99c","vega-driving-language-instructions-en","Vega: Driving with Natural Language Instructions","2026-03-28T14:54:04.698882+00:00",{"id":101,"slug":102,"title":103,"created_at":104},"a34581d6-f36e-46da-88bb-582fb3e7425c","personalizing-autonomous-driving-styles-en","Drive My Way: Personalizing Autonomous Driving Styles","2026-03-28T14:54:26.148181+00:00",{"id":106,"slug":107,"title":108,"created_at":109},"2bc1ad7f-26ce-4f02-9885-803b35fd229d","training-knowledge-bases-writeback-rag-en","Training Knowledge Bases with WriteBack-RAG","2026-03-28T14:54:45.643433+00:00",{"id":111,"slug":112,"title":113,"created_at":114},"71adc507-3c54-4605-bbe2-c966acd6187e","packforcing-long-video-generation-en","PackForcing: Efficient Long-Video Generation Method","2026-03-28T14:55:02.646943+00:00",{"id":116,"slug":117,"title":118,"created_at":119},"675942ef-b9ec-4c5f-a997-381250b6eacb","pixelsmile-facial-expression-editing-en","PixelSmile Framework Enhances Facial Expression Editing","2026-03-28T14:55:20.633463+00:00",{"id":121,"slug":122,"title":123,"created_at":124},"6954fa2b-8b66-4839-884b-e46f89fa1bc3","adaptive-block-scaled-data-types-en","IF4: Smarter 4-Bit Quantization That Adapts to Your Data","2026-03-31T06:00:36.65963+00:00"]