# Open-Source VLA with 36 DOF Dual Arms + Dual Dexterous Hands: 66.7% on Dexterous Tasks, 15 Points Higher than GR00T N1

The noise in teleoperation data has to go into the loss function

Full article: https://haiguangboy.com/posts/dexora

## Core Method

The upper arms use a custom exoskeleton backpack to capture shoulder, elbow, and wrist angles, mapping directly to joint space, while fingers use markerless 3D skeleton retargeting from Apple Vision Pro, avoiding the inverse kinematics jitter and singularities common in pure vision retargeting. The same interface drives both the physical robot and the MuJoCo digital twin, with 20Hz temporal alignment.

## Key Results

It's common knowledge that teleoperation data quality varies, but this work makes it part of the training objective. The approach has three steps: first, pre-filter by kinematic smoothness (RMS of acceleration and jerk), then replay to verify and retain collision-free segments that complete tasks, yielding about a 15% high-quality subset; next, use a simulation-pretrained policy to score each segment on "how well the policy can explain it"; finally, train a discriminator with that 15% as positive samples and the rest as an unlabeled pool, converting the output scores into a weighted diffusion loss.

The ablation cuts both ways: success rates go from 85%→95% and 55%→80%, joint acceleration and jerk drop in tandem, and failures can be traced to specific joints—without the discriminator, the left hand's 5th joint oscillates at high frequency, causing the corn to drop.

## Comparison of Related Approaches

- Opposite route · [Sudu Technology's WAIC debut: Reality is the endpoint for robots, not the boundary of training](https://haiguangboy.com/posts/sudu-third-position) `wx_晚点latepost_20260718_2026_07`: The gap from simulation to dexterity: simulation can guide basic skills, but dexterity must rely on real-world data contradicts the precise claim of a simulation-first training route: pretraining uses almost no real-world data, but after deployment there is failure retry plus a small amount of real-world RL
- Opposite route · [The overlooked non-consensus is whole-body unified control](https://haiguangboy.com/posts/gemini-robotics-2) `deepmind_blog_gemini_robotics_2_brings_20260730_2026_07`: Positioning: existing VLAs are either dual-arm low-DOF grippers or single-arm dexterous hands, not both contradicts dexterity: simultaneously covering 22-DOF five-finger hands and standard two-finger grippers
- Same route · [1,000 tasks in 1 day, relying on inductive bias](https://haiguangboy.com/posts/mt3-thousand-tasks) `learning_a_thousand_tasks_in_a_day_2026_08`: A point masked by numbers: the gap between 66.7% on dexterous tasks and 90%+ on basic tasks remains huge validates perception dependence: vision-only, single camera, no tactile sensing, relies on accurate segmentation
- Same route · [An_Open_Foundation_Model_Towards](https://haiguangboy.com/posts/an_open_foundation_model_towards) `an_open_foundation_model_towards_2026_07`: Three-stage recipe: simulation pretraining → train discriminator → weighted real-world post-training validates ablation confirming the three components contribute independently and additively (pretraining/post-training/RTC)
- Same route · [LAWM: Why action labels become a burden](https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling) `latent_action_pretraining_through_world_modeling_2026_07`: LAWM: Why action labels become a burden
- Same route · [T-Rex: Why tactile sensing needs separate modeling](https://haiguangboy.com/posts/t_rex_tactile_reactive_dexterous_manipulation) `t_rex_tactile_reactive_dexterous_manipulation_2026_07`: T-Rex: Why tactile sensing needs separate modeling

## Limitations

Baseline comparisons are not fully equivalent: π0 requires an external 2-layer MLP projector to raise its output to 36 dimensions, and how much of that 26.7% comes from this temporary adaptation is not distinguished; GR00T N1 uses a 6-DOF setup. The real-world data scale appears in two forms in the text: 177.5 hours versus 40.5 hours. The title touts open-source, but the body lacks a repository URL, license, or release timeline. There is no tactile sensing throughout; the fingertip pads are hard and have low friction, causing slipping, and the authors themselves list tactile sensing as the primary future direction.

## Author's Assessment (Not from the paper, cross-paper synthesis)

The most notable element is the discriminator. It requires no manual quality labeling—using an already-trained policy as a ruler to measure how much each demonstration "resembles actions it recognizes," then converting that score into training weights. This idea is reusable for any system relying on human-collected data, and it is isomorphic to another judgment: when data is scarce, don't have one model learn multiple things at once; anything that can be offloaded should be, and quality filtering counts as one such thing.

Another point is a divergence. This paper concludes that simulation can only guide basic skills, and dexterity must rely on real-world data—pure simulation is near zero on dexterous tasks. The other route bets the opposite: pretraining uses almost no real-world data, relying on failure retry plus a small amount of real-world RL after deployment. Both sides have empirical evidence; the bet is on "whether dexterity can be simulated at all," and the direction should become clear within six months.