frontpage.
newsnewestaskshowjobs

Made with ♥ by @iamnishanth

Open Source @Github

fp.

Open in hackernews

Orbit – Diagnose why your robot policy fails and what data to collect next

https://github.com/Rahillasne/Orbit
1•rahillasne•1h ago

Comments

rahillasne•1h ago
Hi HN, I'm a college sophomore studying CS/Linguistics. I've been building a robot (Handybot) that cleans car interiors using learned policies, and kept running into the same problem: the policy works great in my training setup but fails in real deployments, and I had no way to figure out why. ORBIT is the tool I built to solve this. It logs deployment episodes, detects failures with heuristic detectors (gripper drops, stalls, timeouts), then does something I haven't seen elsewhere — it computes SigLIP vision embeddings for both deployment and training data, measures the distribution gap using FAISS, clusters failure modes with HDBSCAN, and generates ranked prescriptions for what specific demonstrations to collect next. The core insight: most deployment failures happen because the robot encounters visual conditions it wasn't trained on (different lighting, object positions, backgrounds). If you can measure that gap precisely, you can close it with targeted data collection instead of blindly collecting hundreds more demos. Built for HuggingFace's LeRobot framework. Live demo (no install): https://huggingface.co/spaces/Drahils/orbit-demo Limitations I'm upfront about: only works for camera-equipped manipulation arms, can't catch policy architecture bugs or timing issues, and the embedding gap analysis only surfaces visual distribution shifts. Probably covers ~40-50% of real deployment failures — but that's the chunk nobody has tooling for. Would love feedback, especially from anyone deploying learned robot policies in production.

Serve Markdown to LLMs from your Next.js app

https://github.com/kasin-it/next-md-negotiate
1•kasin-it•56m ago•1 comments

Idea Hunting Is Dead. Databases Like This Are Replacing It

1•briebella•58m ago•0 comments

Magawa the HeroRAT

https://en.wikipedia.org/wiki/Magawa
1•KolmogorovComp•1h ago•0 comments

The Lazy Way to Find Your Next SaaS Idea

1•DownCreater•1h ago•1 comments

Pentagon puts Scouts 'on notice' over DEI and girl-centered policies

https://www.npr.org/2026/02/27/nx-s1-5729159/pentagon-scouts-hegseth-dei-deal
1•iamflimflam1•1h ago•0 comments

We Will Be Divided

https://we-are-divided.com
2•shiftybitshiftr•1h ago•0 comments

Social media use/mental wellness links in youth are artifact of other factors [pdf]

https://www.christopherjferguson.com/Brainwaves.pdf
1•mpweiher•1h ago•0 comments

Agentic Engineering Patterns

https://simonwillison.net/guides/agentic-engineering-patterns/
1•vinhnx•1h ago•0 comments

Might Be the Most Underrated Resource for Early Stage Founders

1•vibecoder21•1h ago•0 comments

Kakveda open source-level AI, Infra observability agent

https://www.kakveda.com/
1•prateekdalal•1h ago•1 comments

What Is It That Makes Today's Homes So Different, So Appealing?

https://smarthistory.org/richard-hamilton-just-what-is-it/
1•Torwald•1h ago•0 comments

How LDAP Works

https://growingswe.com/blog/ldap
1•vismit2000•1h ago•0 comments

How you invent math: From counting to complex numbers

https://growingswe.com/blog/inventing-math
1•vismit2000•1h ago•0 comments

Stop Burning Your Context Window – How We Cut MCP Output by 98% in Claude Code

https://mksg.lu/blog/context-mode
1•mksglu•1h ago•1 comments

Zerobyte

https://github.com/nicotsx/zerobyte
1•Erenay09•1h ago•0 comments

Researchers praise 'stunning' results of new prostate cancer treatment

https://www.theguardian.com/society/2026/feb/28/researchers-praise-stunning-results-of-new-prosta...
2•sandebert•1h ago•0 comments

Robotic Land Drone holds positions in Ukraine frontline for 45 days

https://www.youtube.com/watch?v=JMTSH1U44cw
2•lifeisstillgood•1h ago•0 comments

The Missing Piece of AI Discourse Is Trust

https://philipptanlak.com/the-missing-piece-of-ai-discourse-is-trust/
1•philippta•1h ago•0 comments

Modula-2

https://en.wikipedia.org/wiki/Modula-2
3•tosh•1h ago•0 comments

US Customs destroy a rare floppy disk containing demo version of Tsukihime

https://www.timeextension.com/news/2026/02/literally-crying-right-now-50-copies-of-this-adult-onl...
2•FMecha•1h ago•0 comments

ErsatzTV Has Been Archived

https://github.com/ErsatzTV/ErsatzTV/issues/2839
1•kassner•1h ago•0 comments

The Ml.energy Leaderboard

https://ml.energy/leaderboard/
1•kgolubic•1h ago•0 comments

Jeffrey Epstein saw promise in Bitcoin – and its far-right supporters

https://www.theverge.com/tech/885252/jeffrey-epstein-bitcoin-cryptocurrency-connections
1•robtherobber•1h ago•0 comments

Surface-Stable Fractal Dithering

https://runevision.com/tech/dither3d/
1•fanf2•1h ago•0 comments

Show HN: I Turned My CV into a Space Invaders Game (Rust and WASM)

https://breezko.dev
1•breezk0•1h ago•0 comments

Show HN: JobSeek Expat – a CLI tool to scrape English speaking jobs on LinkedIn

https://github.com/electricalgorithm/jobseek-expat
2•gyokhan•1h ago•0 comments

Accuracy vs. Speed in Local LLMs: Finding Your Sweet Spot

https://grigio.org/accuracy-vs-speed-in-local-llms-finding-your-sweet-spot/
1•grigio•1h ago•0 comments

Silent March (2021)

https://tobiaskucera.art/en/tichy-brezen/
1•theobias•1h ago•0 comments

Nano Banana 2 AI Image Generator – Pro Quality, Flash Speed

https://veevid.ai/nano-banana-2
1•jacksteven•1h ago•0 comments

Czinger 21C review: flat out in America's £1.6M, 3D-printed hypercar

https://www.topgear.com/car-reviews/czinger/21c/first-drive
1•teleforce•1h ago•0 comments