Pi robot runs solo for hours at a chocolate factory
Chelsea Finn says a Pi robot ran fully autonomous with no interventions at Dandelion Chocolate, and the hardest part was not the dexterous box assembly but stacking the finished boxes without toppling the pile. Agility also dropped Digit's bird legs, and researchers started poking at what Reward AI means by "no teleop".
Read the whole week in one goFigure's 237 out of 420 became the week's fight
The big one
Pi robot builds and stacks boxes for hours at Dandelion Chocolate
Finn posted a video from a Pi robot deployed at Dandelion Chocolate, which she says is fully autonomous with no interventions. She expected box assembly to be the hard part since it is the most dexterous step, but the real problem was reliably stacking the finished boxes: the original table-mounted robot had poor visibility of the stack, each box lands in a different spot, and one sloppy placement can collapse the pile many boxes later. They moved the deployment to a mobile robot, which she says now runs self sufficient for multiple hours.
A video from a Pi robot deployed at Dandelion Chocolate, fully autonomous w/ no interventions.
Counterintuitively, the hardest part was reliably stacking the boxes.
Stacking was harder than making a box because each step must be completed near-perfectly or the pile will collapse.
Shown this week
Agility's Digit 5 ditches the bird legs, claims 22.7 kg lifts
Agility Robotics unveiled Digit 5, with forward-oriented legs replacing the old ostrich-style backwards knees, swappable end effectors and better battery life. Ziegler's writeup lists repeatable lifts up to 22.7 kg, 90 minutes of work on 9 minutes of charge, 129 kg at 1.81 m tall, and 360 degree person detection with no fixed safety zones, backed by an independent safety controller that can cut motor power. Co-founder Jonathan Hurst is quoted saying the new legs are not about looking human. Early access is H1 2027, general availability by the end of 2027, with first commercial availability outside North America in the EU and UK.
The most important change is the design for safety. Digit 5 can work near people. It stops or kneels when people approach
The previous leg configuration was really optimized for very good walking and running. The new configuration is optimized for lifting heavy things and getting up and down off the ground a lot.
the model decides what's nearby, something verifiable decides when to stop.
Teleop and other arguments
Researchers start picking at Reward AI's "no teleop" claim
Reward AI's OM-1 was announced as zero-shot generalizing to table-top arms, industrial arms and humanoids, learned from human manipulation data with no teleop or robot data. Tripathi called the demo impressive but said the motions do not look like how humans actually move, and Held replied that human data here probably means UMI-style collection, meaning handheld rigs rather than bare hands. Clara Cheng posted three open questions about the action representation, how human data enters the pipeline, and how heavy the per-embodiment integration really is. Paxton pushed back on the "omni embodiment" label, and separately noted that three companies now pursuing human-video learning are all using dexterous gloves for data collection.
No human I know would unplug the cable quite like this. We use opposable thumbs.
I think "human data" includes UMI-style data collection?
it's still one embodiment, it's the hands. turns out all you need are the right hands in the right place
Also on the timeline
Odyssey-3 world model claims it can drive, fly and control robots
Odyssey unveiled Odyssey-3, pitched as a foundation world model that can control robots, power humanoids, drive cars, pilot drones and play video games, with Oliver Cameron saying one researcher taught it to drive the roads of India on 20 hours of driving data. Flexion announced a partnership on top of it the same day, and Animesh Garg said the idea of a single model is starting to emerge.
7 hours of one dataset beat 17 hours of public data in a robot benchmark
Raj Patel posted a benchmark pitting 7 hours of Human Archive data against 17 hours across four widely used public datasets: HOT3D, H2O, HOI4D and TACO. Human Archive alone hit 85% task success and 65.6% task progress. The four public datasets together got 57.5% task success and 52.5% task progress, and mixing all five landed at 60% and 56.3%. The real story: this is one benchmark from one account, not an independent replication, and the sources do not say what the tasks were or what model was trained. What makes it interesting is that adding more data made things worse, which is the opposite of the usual scaling story. Chris Paxton's read is that a lot of public robot data is simply low quality.
Light-Loco-Parkour picks its own way over rough terrain
Light Origins' Light-Loco-Parkour learns a policy that decides when to walk, vault or climb as it moves through a complex environment, using only onboard sensing. RoboPapers says it uses sparse seeds instead of a large motion corpus, and learns when to use its skills without specific sub-task labels. @ChemXiaodao and @Yuntao144 went into the details on episode 104 of RoboPapers with hosts @micoolcho and Chris Paxton. The real story: what is public so far is the episode and a video Chris Paxton calls zero-shot navigation of complex terrain, all on onboard sensing and compute. The sources do not break down how many trials were run or how often it fails, so treat it as a demo plus a conversation with the authors, not a benchmark.
Brad Porter needles the ARR redefinition sweeping robotics
Brad Porter called the pivot from ARR meaning Annual Recurring Revenue to Annualized Run Rate a tremendous sign of the times, then told a reply thread there is a lot of pressure to claim predictability before you have it. Chris Paxton, separately, said it would be underwhelming if the GPT moment for robotics turns out to be just another version of GPT.