INDEPENDENT / UNOFFICIALIndependent and unofficial. Not affiliated with, endorsed by, sponsored by, or operated by Unitree Robotics.
Reporting and analysis on the transition to physical intelligence.
NVIDIA’s policy handles steering, reaching, recovery, teleoperated pickup and sitting through changing goal prompts. Its 83% reaching result and push tests are MuJoCo results, while hardware evidence remains qualitative.
The system combines LiDAR planning, variable-height walking and crawling. Its 48/50 real-robot transition result is measured; the full confined-space runs are still demonstrations.
A foothold predictor reduces the terrain input from 693 heights to 102 values and reaches the training curriculum roughly three times sooner. Real G1 videos show transfer, not measured reliability.
The research system completed 31 of 50 staged real-robot trials with no teleoperation or external tracking. Its five short interactions are not evidence of safe care or workplace deployment.
Attendee footage shows the humanoid falling during a side flip and losing its right lower leg near a joint. Without Unitree’s inspection findings, the video cannot identify the cause or failure rate.
A differentiable world model beats a 128-GPU motion-tracking baseline’s return after three hours on one H100. The headline result is simulated; the physical G1 evidence remains a demonstration.
A diffusion planner lets G1 choose among ducking, sidestepping and turning motions at run time. The physical evidence is promising, but it amounts to five trials per scenario.
A learned controller helps G1 keep a ball or box supported as it changes grip. The tests are human-guided, and the full shelf-to-table sequence succeeds in seven of ten trials.
Alibaba researchers pair 500 hours of human activity with whole-body motion to train a G1 policy. Their real-world trials support the approach, while the release still stops short of a reproducible deployment.
QUT says total programme funding exceeds A$9 million and names partners across care, construction and manufacturing. The announcement sets research scope, not deployment results.
Official code now shows GPT-6 Astra coordinating four G1-D and R1-D skills through cameras, bounded tools and task instructions. The edited demos still provide no trial counts or generalisation benchmark.
Instead of asking one policy to control every joint, Workhorse separates visual planning from whole-body tracking and connects them through five human-scale poses. The result is dynamic autonomy on a real G1—but not yet a general or open system.
A construction-site study got G1 to carry a tool and paint a test sheet, but the clearest result was the productivity gap: about 18 times slower for transport and three times slower for painting.
The September 20 code release makes action-expert training possible with earlier ER weights. It does not yet provide the complete WLA-Base checkpoint or a turnkey reproduction of Unitree’s robot demonstrations.
QUTie’s motion-captured wellness routine will put a humanoid in front of a festival crowd on 22–23 September. The choreography is the easy part; how people choose to approach, follow and trust it is more revealing.
The new G1 EDU teleoperation offering puts integration, data capture and support at the centre of the pitch. Its value depends on the work it saves researchers.
UNITREE AS2
Woodland, shallow water, a carried bag and a racket cameo: what Unitree’s February 2026 As2 reel establishes, and what needs a longer test.
UNITREE G1
A retrospective on Unitree’s 13 May 2024 G1 launch reel: mechanical range, selected hand interactions and the gap between a pose and a reliable task.
10–16 September 2026, reviewed through 14 September. A reading of Unitree’s current G1+ specification highlights sensing, configuration and developer access—the details that matter before an upgrade.
3–9 September 2026 in review. The UnifoLM-X2-1.0 announcement put real-time interaction in the spotlight. The next questions concern the test conditions, supervision and transfer to useful tasks.
Unitree’s first post-listing quarter shows fast revenue growth and much higher R&D spending. Profit, operating cash flow and product-level evidence make the picture less simple.
20–26 August 2026 in review. Wang Xingxing described a threshold for robotics’ breakout. Turning that ambition into a useful test requires more than a percentage.
Its 19 August Shanghai listing is a milestone for the robot maker. Turning public-market attention into useful, dependable machines is a different challenge.