Skip to content
Latest briefingsTopicsCompaniesGlossaryAbout

KAIST humanoid pushes 35 kg and lifts 20 kg while walking, guided by its onboard cameras

KAIST's DRCD Lab released video of Humanoid v0.9 pushing 35 kg objects and lifting 20 kg while walking, using onboard camera feedback rather than external motion capture. The lab credits deep reinforcement learning and human-demonstration learning for the interaction skills.

Robrief Newsroom
2 min readSources verified

KAIST's Dynamic Robot Control & Design Laboratory released a video on Oct 19 titled Whole-Body Locomanipulation with KAIST Humanoid v0.9, in which the machine pushes objects weighing up to 35 kg and lifts up to 20 kg while walking, using onboard camera feedback to arrange the letters K-A-I-S-T. The arms come from a collaborator, RAILAB.

The smaller number deserves the attention. Moving a 35 kg object along a surface never asks the robot to support its weight, and the resistance depends on the contact; a 20 kg lift is a distinct test of what the arms can hold. Quoting the push figure as the payload overstates the machine, and quoting either as a rating is premature, because the lab presents both as demonstrations rather than a commercial specification.

Balance and force are the same problem now

Whole-body locomanipulation means the legs, torso and arms are coordinated against a single balance constraint: pushing through the arms moves the centre of mass, so the legs have to answer for it. That is why a push demonstration is still useful even though it is not a payload test. The v0.9 platform builds on v0.7, a 75 kg, 165 cm machine that ran at 12 km/h and walked on uneven terrain using proprioceptive sensing alone, and whose moonwalking and soccer demos circulated earlier in 2026.

The autonomy claim needs a boundary

KAIST says the interaction skills come from deep reinforcement learning together with learning from human demonstrations. The video description does not detail how the two methods are combined or how a task is commanded, so onboard visual feedback on its own should not be read as unrestricted autonomy. Tech Times frames the 35 kg push as roughly double the nearest published result for a humanoid manipulating from onboard vision, which is a comparison worth keeping at the level of published figures.

What v0.9 changes is that a research platform known for agile walking now does visually guided physical work without a motion-capture rig. Whether that holds up is a question for repeated tasks on a factory floor, which is the step no demonstration video settles.

Sources Sources verified

  1. 1.Humanoids Daily: From Moonwalking to Heavy Lifting: KAIST Humanoid v0.9 Moves 35 kg Objects (Oct 19, 2026)humanoidsdaily.com
  2. 2.Tech Times: KAIST Humanoid Pushes Objects Up to 77 Lbs While Walking, No External Cameras (Oct 20, 2026)techtimes.com
  3. 3.Gadget Review: KAIST Humanoid v0.9 Robot Pushes 77 Pounds and Lifts 44 While Walkinggadgetreview.com

Dyna's Lindon Gao says one working workflow equals all of them, and has moved spending toward pretraining

Dyna Robotics founder and CEO Lindon Gao told Zhang Xiaojun on episode 154 of the podcast 商业访谈录 that pretraining now takes much more of the company's resources than post-training, up from an even split. Dyna's Sept 29 post for Dyna-2.1 says a laundry cycle chains about 79 steps and that at 95% per step it almost never finishes without help, and the machine it showed, Taku, ran an uncut autonomous one-hour laundry-room workflow.

5 min readSources verified

Gemini 4 Argon scores 87% on embodied reasoning but takes 52 minutes per real-arm task

Google made Gemini 4 Argon globally available on Oct 13 and grouped its robotics uses into robot control, Real2Sim, physical understanding, and research and self-improvement. Artificial Analysis's AA-Robotics v0.1 scored Argon 55.3 out of 100 driving a real arm on eight tabletop tasks, about 100 times slower than a person, while Physical Intelligence opened an API beta the same day for fine-tuning its π0 and π0.5 models on Pi's infrastructure.

4 min readSources verified

LoopWAM gains 19.7 points on real robots by retraining on its own labeled rollouts

LoopWAM is a closed-loop world-action model from Tsinghua University, Dexmal and Loopeva that scores every rollout by quality and gives failed trajectories video supervision but no action supervision. It ranks first on RoboChallenge's Table30-V2 at 43.33% success, 2.67 points over the VLA-DM0.5 entry in its own table and 0.33 points over the 43.0% Dexmal publishes for DM0.5.

5 min readSources verified

Newsletter

Robrief Weekly

One email a week with the robotics news that matters. Free, unsubscribe anytime.