PaperDSD
Direction-Scale Decomposition in Action Representation: Rethinking What to Tokenize for Vision-Language-Action Models
Claim tested: DSD's direction–scale action representation improves real-robot task success over standard action binning, without additional robotics pretraining.
Tested
Result
Improvement over baseline
Replicated
- Reported
- +13 pts
- Ours
- +13 pts
- Gain vs baseline
- 95% CI, +2 pts to +24 pts
- Paper's reported gain
- No difference
Success rate
- Reported
- 73%
- Ours
- 45%
- Measured on our stations
- 95% CI, 38%–53%
- Reported in the paper
Across iterations
By task
| Task | BaselineUniform action binning | MethodDSD | Improvement |
|---|---|---|---|
| e01Clamp electrical component | 10%3/30 | 3%1/30 | −7 pts−23 to +8 |
| e02Open the toolbox with a screwdriver | 33%10/30 | 20%6/30 | −13 pts−34 to +9 |
| e03Place loaded boxes in container (bimanual) | 30%9/30 | 60%18/30 | +30 pts+5 to +50 |
| e04Pour screws into the funnel | 60%18/30 | 60%18/30 | 0 pts−23 to +23 |
| e05Datum-corner alignment | 27%8/30 | 83%25/30 | +57 pts+32 to +72 |
| All tasks | 32%48/150 | 45%68/150 | +13 pts+2 to +24 |
| Reported in the paper | 60% | 73% | +13 pts |
Rollouts
One clip per environment from this entry's evaluation, beside where each evaluation episode placed the environment's objects, coloured by its rating. Use the button beside each environment to open its evaluation dataset in the dataset viewer.
e02Open the toolbox with a screwdriver BaselineUniform action binning10/3033%
Show in dataset viewer of Uniform action binning on Open the toolbox with a screwdrivere04Pour screws into the funnel BaselineUniform action binning18/3060%
Show in dataset viewer of Uniform action binning on Pour screws into the funnelMethodDSD18/3060%
Show in dataset viewer of DSD on Pour screws into the funnel- Held-out benchmark environment
e01Clamp electrical component BaselineUniform action binning3/3010%
Show in dataset viewer of Uniform action binning on Clamp electrical componentMethodDSD1/303%
Show in dataset viewer of DSD on Clamp electrical component - Held-out benchmark environment
e03Place loaded boxes in container (bimanual) BaselineUniform action binning9/3030%
Show in dataset viewer of Uniform action binning on Place loaded boxes in container (bimanual) - Held-out benchmark environment
e05Datum-corner alignment BaselineUniform action binning8/3027%
Show in dataset viewer of Uniform action binning on Datum-corner alignmentMethodDSD25/3083%
Show in dataset viewer of DSD on Datum-corner alignment
Implementation Details
What we changed
We compare DSD with standard action binning to test whether separating motion direction from magnitude improves real-robot task success without additional robotics pretraining. We reimplemented both representations on the same Florence-2 policy, starting from its public weights and following the paper’s parallel action decoding. The selected BIN baseline uses the paper’s percentile clipping. Both arms train for 40,000 steps at batch 32 on one GPU.
- Longer chunks. We predict 30 steps instead of 20 because controls run at 30 Hz, giving a one-second action chunk.
- Absolute-command execution. Predicted relative increments are accumulated into absolute commands. The first increment is measured from the observed pose and has separate normalization statistics; later increments are differences between successive commands. Both arms use this construction.
- Exact-zero decoding. Motion bins containing zero decode to exactly zero instead of their bin centers, preventing idle drift when increments are accumulated. This replaces reliance on a controller dead zone and applies to both arms; grippers are unchanged.
- DSD scale bounds. BIN uses the paper’s 1st/99th-percentile clipping on every coordinate, while DSD maps its magnitudes from zero to their 99.9th percentile and keeps fixed bounds for directions. The paper does not fully specify DSD’s scale percentile; this choice retains more fast-motion targets and may affect the comparison.
- Direction-loss weighting. Our DSD weights the direction loss with a soft ramp,
w = clip(s/τ, 0.1, 1), with τ = 0.3% of the full-scale range. The authors’ implementation uses a step: weight 0.1 below the threshold and 1.0 above it. In the authors’ LIBERO experiments, switching to the ramp alone reduced success by about 3 points, bringing DSD close to their binning baseline. The authors also recommend a lower threshold for our action definition (consecutive pose differences): 0.05–0.1% of the full-scale range, compared with the 0.3% we used. Both differences may have reduced DSD’s performance in our evaluation; their effect on our tasks was not measured.
Prompt summary
Use the paper's 1/99 percentile bounds with clipping for the BIN arm, retrain and re-evaluate the five BIN models, and update the comparison. Reuse the DSD models and evals, pair each new BIN eval with its earlier placements, and compare the 1/99 BIN against the earlier 0.1/99.9 BIN.
Response summary
DSD 45.3% vs BIN 1/99 32.0% SR (+13.3 pp, 95% CI +4.0 to +22.7, p = 0.02; AP +8.6 pp, p = 0.07), driven by push (25/30 vs 8/30) and boxes (18/30 vs 9/30). The 0.1/99.9 BIN beat the 1/99 BIN by +14.0 pp.
Conclusion
DSD outperformed the paper's binning baseline (1st/99th-percentile clipping) on our robots: 45% vs 32% success (+13 pts, 95% CI +2 to +24). In an additional comparison, a binning baseline with wider bounds (0.1/99.9 percentiles) performed on par with DSD in our setup. We report this as a single data point under our controller and action definition. It does not establish that clipping explains DSD's gains, particularly as our DSD weights its direction loss differently from the authors' implementation, and the authors observed no comparable effect from widening the bounds in their setup. Scope: this evaluation covers a single-embodiment setting without additional robotics pretraining. It does not test the paper's other claims, including mixed-dataset training across embodiments.