[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-factr-2-force-sensing-robot-arms-en":3,"article-related-factr-2-force-sensing-robot-arms-en":30,"series-research-f20df85e-3b45-4eec-a44c-7fa0940e0d39":81},{"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":29},"f20df85e-3b45-4eec-a44c-7fa0940e0d39","factr-2-force-sensing-robot-arms-en","FACTR 2 brings force sensing to cheap robot arms","\u003Cp data-speakable=\"summary\">FACTR 2 adds force sensing to commodity robot arms and uses it to improve policy learning.\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>Research org\u003C\u002Fstrong>: Unspecified in arXiv abstract\u003C\u002Fli>\u003Cli>\u003Cstrong>Core data\u003C\u002Fstrong>: 10 minutes of free-motion data\u003C\u002Fli>\u003Cli>\u003Cstrong>Breakthrough\u003C\u002Fstrong>: Neural External Torque Estimation from free-motion data\u003C\u002Fli>\u003C\u002Ful>\u003Cp>Contact-rich robot work is hard when the arm cannot sense force. That gap matters for anyone building teleoperation systems, imitation learning pipelines, or manipulation policies on low-cost hardware, because touch is often the difference between a motion that looks correct and one that actually succeeds.\u003C\u002Fp>\u003Cp>This paper argues that you do not need dedicated force sensors to get useful force awareness. Instead, the authors introduce a learned estimator for external joint torque, then use those estimates to make both teleoperation and policy training more contact-aware on commodity robot arms.\u003C\u002Fp>\u003Ch2>What problem this paper is trying to fix\u003C\u002Fh2>\u003Cp>Many robot arms ship without dedicated force sensors, largely because those sensors add cost. That is a real constraint for labs, startups, and robotics teams trying to scale manipulation beyond premium platforms. Without force feedback, it is harder to detect pre-contact, contact, and collision phases, all of which matter in tasks that involve grasping, insertion, pushing, or other contact-rich behaviors.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781159586943-1hon.png\" alt=\"FACTR 2 brings force sensing to cheap robot arms\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>The paper focuses on two downstream problems caused by that missing signal. First, teleoperation becomes less informative because the operator cannot feel or infer contact as directly. Second, policy learning from demonstrations can miss the parts of trajectories where force is most important, which can leave a learned policy brittle exactly when it needs to be precise.\u003C\u002Fp>\u003Cp>In other words, the paper is not just about sensing for its own sake. It is about making low-cost robot arms more usable for learning and control in the places where touch matters most.\u003C\u002Fp>\u003Ch2>How the method works in plain English\u003C\u002Fh2>\u003Cp>The first piece is Neural External Torque Estimation, or NEXT. The abstract describes NEXT as a data-driven method that estimates external joint torques without any dedicated force sensors. The key practical detail is that it trains from only 10 minutes of free-motion data and does so in 1 minute, which suggests the method is designed to be lightweight enough to fit into real robot workflows.\u003C\u002Fp>\u003Cp>Free-motion data is important here because it means the robot does not need a special force-sensing setup or a complicated calibration routine to get started. The model learns from how the arm moves on its own, then infers external torque from that motion. The abstract says the resulting estimates are comparable to dedicated joint-torque sensors, which is the core claim that makes the approach interesting.\u003C\u002Fp>\u003Cp>The second piece is Force-Informed Re-Sampling Training, or FIRST. This is the policy-learning side of the system. Rather than treating all demonstration frames equally, FIRST up-samples pre-contact and contact segments during behavior cloning. That means the training process pays more attention to the moments where the robot is about to touch something or is already in contact, which are often the most informative and the most failure-prone parts of a trajectory.\u003C\u002Fp>\u003Cp>Together, NEXT and FIRST form a pipeline: estimate force signals on a cheap arm, then use those signals to bias learning toward the contact-rich parts of the data. That is a simple idea, but it is the kind of simple idea that can change how well a manipulation policy learns from the same demonstrations.\u003C\u002Fp>\u003Ch2>What the paper actually shows\u003C\u002Fh2>\u003Cp>The abstract gives three concrete results. First, NEXT trains in 1 minute from 10 minutes of free-motion data. Second, it achieves estimates comparable to dedicated joint-torque sensors. Third, it enables force-feedback teleoperation on low-cost arms and improves policy learning through FIRST.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781159595696-wkkh.png\" alt=\"FACTR 2 brings force sensing to cheap robot arms\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>On the policy side, the paper says that across five long-horizon tasks, FIRST outperforms prior force-aware policies by over 17% in task progress. That is the only \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa>-style number provided in the abstract, so there are no additional task names, success rates, or hardware details to lean on here. Still, the direction is clear: the force-aware sampling strategy appears to make imitation learning better on tasks where contact matters over long horizons.\u003C\u002Fp>\u003Cp>It is also worth separating what is shown from what is implied. The abstract does not claim that NEXT replaces all force sensors in every setting, and it does not give a full breakdown of error metrics for torque estimation. It also does not spell out how the “comparable” claim was measured. So the safe reading is that the method looks promising and practical, but the abstract alone does not provide enough detail to judge its limits across robot models, payloads, or environments.\u003C\u002Fp>\u003Cul>\u003Cli>NEXT estimates external joint torque without dedicated force hardware.\u003C\u002Fli>\u003Cli>FIRST oversamples pre-contact and contact segments during behavior cloning.\u003C\u002Fli>\u003Cli>The reported policy gain is over 17% in task progress across five tasks.\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>Why developers and robotics teams should care\u003C\u002Fh2>\u003Cp>If you work on robot learning, the biggest takeaway is that force awareness may be more accessible than the hardware bill suggests. A lot of manipulation stacks are built around the sensors you can afford, not the sensors you wish you had. This paper points to a path where commodity arms can still support force-aware teleoperation and better training data selection.\u003C\u002Fp>\u003Cp>That matters for engineering because it can simplify system design. Instead of integrating dedicated force sensors into every arm, you may be able to estimate useful torque signals from motion data and then feed those signals into your training pipeline. For teams iterating quickly, that could reduce setup complexity and make contact-rich tasks more practical to prototype.\u003C\u002Fp>\u003Cp>It also suggests a useful lesson for imitation learning: not all frames in a demonstration are equally valuable. If contact is where policies fail, then methods like FIRST that overweight those regions may help more than generic resampling tricks. That is especially relevant for long-horizon tasks, where a small mistake near contact can cascade into failure later in the sequence.\u003C\u002Fp>\u003Cp>At the same time, the paper leaves open some obvious questions. How well does NEXT transfer across different robot arms, speeds, payloads, or wear conditions? How sensitive is FIRST to noisy torque estimates? And how much of the reported gain comes from the sensing estimate itself versus the re-sampling strategy built on top of it? The abstract does not answer those questions, so practitioners should treat the result as a promising starting point rather than a finished recipe.\u003C\u002Fp>\u003Ch2>The practical read\u003C\u002Fh2>\u003Cp>The cleanest way to think about FACTR 2 is that it tries to make touch a software problem before it becomes a hardware problem. NEXT extracts a force-like signal from motion, and FIRST uses that signal to teach policies where contact matters most. For developers building on off-the-shelf robots, that combination could be more useful than a more expensive arm with a richer sensor suite, especially if the goal is to scale manipulation learning without redesigning the robot.\u003C\u002Fp>\u003Cp>The paper’s broader message is straightforward: if you can infer contact well enough, you can train better policies and operate cheaper robots more effectively. The abstract suggests that this is already working across multiple long-horizon tasks, but the detailed behavior of the method will matter for real deployments.\u003C\u002Fp>\u003Cp>For now, the main engineering value is the direction of travel. Force sensing does not have to be a hard boundary between low-cost hardware and contact-rich manipulation. This paper argues that learned torque estimation plus force-aware sampling can close part of that gap.\u003C\u002Fp>","FACTR 2 adds force sensing to commodity robot arms and uses it to improve policy learning.","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.12406",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781159586943-1hon.png","research","en","d29d34a6-1c00-4ac8-a751-292abacada4d",[17,18,19,20,21],"robot manipulation","force sensing","behavior cloning","teleoperation","torque estimation",[23,24,25],"NEXT estimates external joint torque from free-motion data without dedicated force sensors.","FIRST up-samples pre-contact and contact segments to improve behavior cloning.","Across five long-horizon tasks, FIRST beats prior force-aware policies by over 17% in task progress.",0,"2026-06-11T06:32:36.958637+00:00","2026-06-11T06:32:36.948+00:00","3103988e-c4fe-45e3-98ab-846500c9d507",{"tags":31,"relatedLang":41,"relatedPosts":45},[32,34,36,38,39],{"name":21,"slug":33},"torque-estimation",{"name":19,"slug":35},"behavior-cloning",{"name":18,"slug":37},"force-sensing",{"name":20,"slug":20},{"name":17,"slug":40},"robot-manipulation",{"id":15,"slug":42,"title":43,"language":44},"factr-2-force-sensing-robot-arms-zh","FACTR 2 讓便宜機械臂感知力道","zh",[46,52,58,64,70,76],{"id":47,"slug":48,"title":49,"cover_image":50,"image_url":50,"created_at":51,"category":13},"3d767fcf-58cf-4a92-9652-00f1ec1e3677","c-dic-incremental-compression-dialogue-memory-en","C-DIC compresses long dialogue memory turn by turn","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781158686061-b943.png","2026-06-11T06:17:38.690213+00:00",{"id":53,"slug":54,"title":55,"cover_image":56,"image_url":56,"created_at":57,"category":13},"e9cb5863-f541-4d53-8f38-289660919a1f","reroute-keeps-useful-vision-tokens-alive-en","Reroute Keeps Useful Vision Tokens Alive","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781157784473-28u1.png","2026-06-11T06:02:32.556043+00:00",{"id":59,"slug":60,"title":61,"cover_image":62,"image_url":62,"created_at":63,"category":13},"ffb2e7ac-bff8-4c03-a4d4-1c19264c6967","sequential-fine-tuning-essay-scoring-en","Sequential fine-tuning improves essay scoring","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781146976434-ism9.png","2026-06-11T03:02:29.860107+00:00",{"id":65,"slug":66,"title":67,"cover_image":68,"image_url":68,"created_at":69,"category":13},"21a693ca-7c72-49e6-886e-4d190baa33c1","nvidia-nemotron-3-ultra-open-models-compete-en","NVIDIA Nemotron 3 Ultra proves open models can still compete","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781108276690-iat5.png","2026-06-10T16:17:24.880013+00:00",{"id":71,"slug":72,"title":73,"cover_image":74,"image_url":74,"created_at":75,"category":13},"3e763732-2a73-4539-8990-b8af7d671b3e","speechllm-l2-assessment-rationales-en","SpeechLLM Gives L2 Scores and Rationales","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781103780733-r41k.png","2026-06-10T15:02:34.030726+00:00",{"id":77,"slug":78,"title":79,"cover_image":11,"image_url":11,"created_at":80,"category":13},"ef3677ab-2c91-4c09-9c61-b19dbd7d12fb","eevee-test-time-prompt-learning-real-world-en","EEVEE tackles prompt learning across real-world streams","2026-06-10T06:32:32.554039+00:00",[82,87,92,97,102,107,112,117,122,127],{"id":83,"slug":84,"title":85,"created_at":86},"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":88,"slug":89,"title":90,"created_at":91},"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":93,"slug":94,"title":95,"created_at":96},"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":98,"slug":99,"title":100,"created_at":101},"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":103,"slug":104,"title":105,"created_at":106},"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":108,"slug":109,"title":110,"created_at":111},"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":113,"slug":114,"title":115,"created_at":116},"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":118,"slug":119,"title":120,"created_at":121},"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":123,"slug":124,"title":125,"created_at":126},"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":128,"slug":129,"title":130,"created_at":131},"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"]