(8/N) Provided Baselines
We include a set of baselines to kickstart your experiments:
• Classic behavioral cloning models: Diffusion Policy, WB-VIMA, ACT, BC-RNN • Pre-trained VLA models: OpenVLA, π_0 @physical_int
MACHINE LEARNING
-
Baseline Models for Robot Learning Experiments
By
–
-
Diverse State Transitions and Manipulation Skills in Robotics
By
–
(7/N) 🧩Feature #4: Diverse State Transitions and Manipulation Skills
— Fei-Fei Li (@drfeifei) 2 septembre 2025
• Spatial: next_to, inside, on_top, under, touching
• Particles: covered, uncovered
• Thermal: hot, cooked, on_fire, frozen
• Others: open, closed, on, off, attached, sliced, diced pic.twitter.com/AS8KBroBxZ(7/N) Feature #4: Diverse State Transitions and Manipulation Skills
• Spatial: next_to, inside, on_top, under, touching
• Particles: covered, uncovered
• Thermal: hot, cooked, on_fire, frozen
• Others: open, closed, on, off, attached, sliced, diced -
Long-Horizon Mobile Manipulation in Realistic Home Environments
By
–
(6/N) 🏠Feature #3: Long-Horizon Mobile Manipulation in Realistic Homes
— Fei-Fei Li (@drfeifei) 2 septembre 2025
• Task durations range from 1 to 25 minutes (average 6.6 minutes)
• Performed in household-scale scenes
• Requires memory, planning, and reasoning over a long period of time pic.twitter.com/LD2ObgdsxT(6/N) Feature #3: Long-Horizon Mobile Manipulation in Realistic Homes
• Task durations range from 1 to 25 minutes (average 6.6 minutes)
• Performed in household-scale scenes
• Requires memory, planning, and reasoning over a long period of time -
High-Quality Teleoperated Data for Robot Training Systems
By
–
(5/N) 💯Feature #2: High-Quality Data
— Fei-Fei Li (@drfeifei) 2 septembre 2025
✔️Teleoperated using our JoyLo interface
✔️Near-optimal, clean demos
✔️Consistent manipulation behaviors
✔️Moderate, consistent teleoperation speed
🚫 No sudden accelerations/decelerations
🚫 No failed grasps
🚫 No unintended collisions pic.twitter.com/kLNUdwYAE2(5/N) Feature #2: High-Quality Data
Teleoperated using our JoyLo interface
Near-optimal, clean demos
Consistent manipulation behaviors
Moderate, consistent teleoperation speed No sudden accelerations/decelerations No failed grasps No unintended collisions -
Large-Scale Robot Demonstration Dataset with 50 Tasks
By
–
(4/N) 🔍Feature #1: Large-Scale Demonstration Dataset
— Fei-Fei Li (@drfeifei) 2 septembre 2025
• 50 tasks, 10,000 demos, a total of ~1,200 hours of data
• Subtask and skill (30+) segmentation
• Spatial relation annotation
• Multi-granularity language annotation pic.twitter.com/XZDl3aptuU(4/N) Feature #1: Large-Scale Demonstration Dataset
• 50 tasks, 10,000 demos, a total of ~1,200 hours of data
• Subtask and skill (30+) segmentation
• Spatial relation annotation
• Multi-granularity language annotation -
Stanford Mobile Manipulation Challenge: 50 Tasks with 1200 Hours Demos
By
–
(2/N) In this challenge, take on 50 long-horizon mobile manipulation tasks that require diverse and complex low-level skills, powered by 1,200 hours of high-quality demonstrations. Challenge website: https://
behavior.stanford.edu -
BEHAVIOR Challenge 2025: Advanced Robot Task-Solving Competition
By
–
(1/N) How close are we to enabling robots to solve the long-horizon, complex tasks that matter in everyday life?
— Fei-Fei Li (@drfeifei) 2 septembre 2025
🚨 We are thrilled to invite you to join the 1st BEHAVIOR Challenge @NeurIPS 2025, submission deadline: 11/15.
🏆 Prizes:
🥇 $1,000
🥈 $500
🥉 $300 pic.twitter.com/W63PyzBcMk(1/N) How close are we to enabling robots to solve the long-horizon, complex tasks that matter in everyday life? We are thrilled to invite you to join the 1st BEHAVIOR Challenge @NeurIPS 2025, submission deadline: 11/15. Prizes: $1,000 $500 $300
-
Engineers and researchers advance Pareto optimization frontiers
By
–
Engineers find a new point on the Pareto curve. Researchers discover a better Pareto curve.
-
OpenAI for Science: AI Revolution in Biology and Scientific Research
By
–
OpenAI for science is doing exactly what Dario Amodei pointed out to us about a year ago, namely that AI will revolutionize the entire field of science, and biology in particular: "If our core hypothesis about AI progress is correct, then the right way to think of AI is not as a
-
Time100 AI List: Key Voices Shaping Artificial Intelligence
By
–
I was recently named to the Time100 AI list – check out the list, not just from 2025 but from all three years for more voices and predictions in AI. From my own list: @emollick @SashaMTL @simonw @conorgrennan @sineadbovell @OfficialLoganK @karpathy @DarioAmodei @ilyasut @sama